Question and answer calculation method, server, and storage medium

Abstract

Embodiments of the present application provide a question and answer computing method, a server and a storage medium. In the embodiments, when automatic question and answer computing is performed for a user question initiated by an object, an answer encoding feature can be calculated according to the user question and description information of the object, and a user role preference feature can be calculated according to the user question and historical data of the user, and then automatic question and answer computing can be performed by comprehensively considering data in multiple different dimensions to match an answer that meets the personalized needs of the user. Based on this implementation, the description information of the object and the historical data of the user can be fully utilized, the effectiveness of the automatic question and answer computing result is improved, and the actual questioning needs of the user can be met.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a question-answering computation method, server, and storage medium. Background Technology

[0002] With the development of artificial intelligence technology, intelligent question answering is gradually entering people's lives. Users can submit questions to question answering platforms through their devices, and the platforms can match suitable answers to the user's questions and return them to the user. For example, in e-commerce scenarios, to provide a better online shopping experience, e-commerce platforms offer Product Question Answering (PQA) services. PQA services can automatically generate product-related answers and act as intelligent online shopping assistants to improve the customer's shopping experience.

[0003] In existing technologies, question-answering platforms typically identify user opinions from relevant comments on a given question to provide a "yes" or "no" answer, or provide an answer based on the semantic relevance between a given question and comment fragments. However, this question-answering computation method does not consider the user's actual needs and cannot provide effective answers to user questions. Therefore, a solution is needed. Summary of the Invention

[0004] This application provides a question-and-answer calculation method, a server, and a storage medium in various aspects to improve the effectiveness of question-and-answer calculation results and meet users' real question-and-answer calculation needs.

[0005] This application provides a question-and-answer calculation method, comprising: receiving a user question sent by a terminal device, wherein the user question is initiated by a user for a target; calculating the answer encoding feature corresponding to the user question based on the user question and the description information of the target; calculating the user's role preference feature based on the user's historical data; and calculating the personalized answer corresponding to the user question using the answer encoding feature and the role preference feature.

[0006] This application embodiment also provides a question-and-answer calculation method, including: displaying a first interface, the first interface including at least one question entry for an object; responding to a question entry triggering operation issued by a user, displaying a question-and-answer interface adapted to the triggered question entry; obtaining a user question initiated by the user for the object through the question-and-answer interface, and sending the user question to a server, so that the server can determine a personalized answer to the user question based on the user question, the description information of the object, and the user's historical data.

[0007] This application also provides a server, including: a memory, a processor, and a communication component; the memory is used to store one or more computer instructions; the processor is used to execute one or more computer instructions to perform the question-and-answer calculation method provided in this application.

[0008] This application also provides a server, including: a memory, a processor, and a communication component; the memory is used to store one or more computer instructions; the processor is used to execute one or more computer instructions to perform the question-and-answer calculation method provided in this application.

[0009] This application also provides a terminal device, including: a memory, a processor, and a communication component; the memory is used to store one or more computer instructions; the processor is used to execute one or more computer instructions to perform the question-and-answer calculation method provided in this application.

[0010] This application also provides a computer-readable storage medium storing a computer program, which, when executed, can implement the steps in the method provided in this application.

[0011] In this embodiment, when performing automatic question-and-answer calculation for user questions initiated by an object, answer encoding features can be calculated based on the user question and the object's description information, and user role preference features can be calculated based on the user question and the user's historical data. This allows for comprehensive consideration of data from multiple dimensions to perform automatic question-and-answer calculation, thereby matching answers that meet the user's personalized needs. Based on this implementation, the object's description information and the user's historical data can be fully utilized, improving the effectiveness of the automatic question-and-answer calculation results and better meeting the user's actual question-and-answer needs. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0013] Figure 1 A schematic diagram of the structure of a question-answering computing system provided for an exemplary embodiment of this application;

[0014] Figure 2 A schematic diagram of the structure of a personalized answer generation model provided in an exemplary embodiment of this application;

[0015] Figure 3 A flowchart of a question-answering calculation method provided as an exemplary embodiment of this application;

[0016] Figure 4A flowchart of a question-answering calculation method provided as another exemplary embodiment of this application;

[0017] Figure 5 A schematic diagram of the structure of a server is provided for another exemplary embodiment of this application;

[0018] Figure 6 This is a schematic diagram of the structure of a terminal device provided for an exemplary embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In some application scenarios, there is a need for question-answering computation. Current technologies typically meet this need by matching user questions with existing questions, finding the matching existing questions, and returning the corresponding answers to the users. Alternatively, they might identify user comments from related comments to provide a "yes" or "no" answer. Or, they might provide an answer based on the semantic relevance between a given question and comment fragments. However, these question-answering computation methods often result in low answer effectiveness. For example, regarding the same product, User 1 might prefer price information, User 2 might prefer quality information, and User 3 might prefer weight information. If a uniform answer is provided without considering the users' actual needs, users will not obtain effective information from the answer.

[0021] To address the aforementioned technical problems, this application provides a solution through some exemplary embodiments, which will be described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic diagram of the structure of a question-answering calculation system provided for an exemplary embodiment of this application. For example... Figure 1 As shown, the question-and-answer computing system 10 includes: a terminal device 11 and a server 12.

[0023] The terminal device 11 and the server 12 can be connected via a wireless or wired network. In this embodiment, if the terminal device 11 communicates with the server 12 through a mobile network, the network standard of the mobile network can be any one of 2G (GSM), 2.5G (GPRS), 3G (WCDMA, TD-SCDMA, CDMA2000, UTMS), 4G (LTE), 4G+ (LTE+), 5G, WiMax, etc.

[0024] Terminal device 11 can be a smartphone, tablet, personal computer, wearable device, etc. There can be one or more terminal devices 11. Terminal device 11 typically includes at least one processing unit, at least one memory, and a display. The number of processing units and memory depends on the configuration and type of terminal device 11.

[0025] The display may include a screen, primarily used to display various types of information. Optionally, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensor can sense not only the boundaries of the touch or swipe action but also information such as the duration and pressure associated with the touch or swipe operation.

[0026] Memory can include volatile components such as RAM, non-volatile components such as read-only memory (ROM) and flash memory, or both. Memory typically stores the operating system (OS), one or more application programs, and may also store program data.

[0027] In addition to the processing unit, memory, and display, the terminal device 11 also includes some basic configurations, such as a network interface card (NIC) chip, an I / O bus, and audio / video components. Optionally, the terminal device 11 may also include some peripheral devices, such as a keyboard, mouse, pen, and printer. These peripheral devices are well known in the art and will not be described in detail here.

[0028] In this embodiment, the terminal device 11 is mainly used to acquire user questions raised by the user regarding a target during human-computer interaction and send the user questions to the server 12. Depending on the human-computer interaction scenario and the implementation of the terminal device 11, the way the terminal device 11 interacts with the user will also vary. Typically, the terminal device 11 can provide users with various forms of human-computer interaction interfaces, such as a human-computer interaction interface and a voice input interface. When a user needs to ask a question, they can input the user question into the terminal device 11 through the human-computer interaction interface provided by the terminal device 11.

[0029] In some application scenarios, the terminal device 11 can provide a first interface to the user as a human-computer interaction interface. This human-computer interaction interface can be implemented as a webpage, an application page, or a window, etc. To facilitate users in asking questions for different categories of problems, the first interface includes at least one question entry corresponding to the question category.

[0030] For example, on a product page provided by an e-commerce platform, a question-asking entry point can be displayed. Users can trigger at least one of these entry points to initiate a question request based on their needs. In response to the user's triggering of the question entry point, terminal device 11 displays a Q&A interface adapted to the triggered entry point. Users can enter their questions (referred to as user questions) on this interface. After receiving the user question, terminal device 11 sends it to server 12, where server 12 matches an answer and returns it. After receiving the answer to the user question from server 12, terminal device 11 can output the answer to the Q&A interface for the user to view.

[0031] In the question-answering computing system 10, server 12 primarily provides question-answering computing services to terminal device 11. For example, it receives user questions sent by terminal device 11, matches answers to those questions, and returns the matched answers to terminal device 11. Server 12 can be any device that can provide computing services and respond to service requests and perform processing, such as a conventional server, cloud server, cloud host, virtual data center, etc. The main components of server 12 include a processor, hard disk, memory, system bus, etc., similar to a general computer architecture.

[0032] In this embodiment, after receiving a user question sent by the terminal device 11, the server 12 can determine the object associated with the user question, at least one associated fact of the object, and the historical data of the user who initiated the user question. When calculating the answer to the user question, the server 12 can calculate the answer encoding features corresponding to the user question based on the user question and the description information of the object, and calculate the user's role preference features based on the user's historical data.

[0033] The object can be implemented as a physical object, such as tangible goods or items; or as a virtual object, such as e-books, audio and video, video games, or network software; or as a service object, such as instant delivery services, food delivery services, housekeeping services, advertising services, or cloud computing services. This embodiment includes, but is not limited to, these.

[0034] The descriptive information of an object, also known as supporting facts, typically includes at least one of the following: product specifications, attribute information, and user reviews. When describing an object, it can be presented from multiple aspects; therefore, the descriptive information can include fragments of information from various aspects. For example, for a camera, the descriptive information fragments could include: lens parameters, brand, weight, focal length, and other aspects. For a smartphone, the descriptive information fragments could include: camera parameters, sound quality, specifications, price, and other aspects.

[0035] Here, user historical data typically refers to user-generated content (UGC), including but not limited to comments, questions, and responses to other users' questions posted by users within a historical time period. This embodiment does not impose any limitations. Based on user historical data, it is possible to analyze a user's preferences regarding one or more aspects of an object. For example, based on a large amount of user 1's historical data regarding object weight, it can be analyzed that user 1 is interested in the object's weight; based on a large amount of user 2's historical data regarding object price, it can be analyzed that user 2 is more interested in the object's price, and so on.

[0036] In this embodiment, server 12 can obtain a large amount of historical data of users and descriptive information of objects. After obtaining a user's question, server 12 can retrieve historical data and descriptive information of objects associated with the user's question from the large amount of historical data and descriptive information based on an information retrieval (IR) system. The historical data and descriptive information described in the following embodiments refer to the historical data and descriptive information of users associated with the question retrieved by the IR system, and will not be described again hereafter.

[0037] In this embodiment, after obtaining the text corresponding to historical data or the text corresponding to descriptive information, the text can be segmented to obtain text fragments. For example, the text corresponding to historical data or the text corresponding to descriptive information can be divided into fragments of a set length; or, according to the length of the sentences, the text corresponding to historical data or the text corresponding to descriptive information can be divided into multiple sentences, with each sentence corresponding to a fragment.

[0038] The data segments obtained by segmenting the descriptive information of the object can be expressed as: F = {f1, f2, ..., fk}.

[0039] The data segments obtained by segmenting the user's historical data can be expressed as: H = {h1, h2, ..., hk}.

[0040] The operation of server 12 calculating the answer encoding features corresponding to the user's question based on the user's question and the object's description information can be implemented based on a text feature extraction algorithm. For example, the word2vec algorithm, GloVe algorithm, fasttext algorithm, or feature extraction algorithm based on an attention mechanism are not limited in this embodiment. The answer encoding features corresponding to the user's question are not affected by the user's personalized information and can be used to calculate answers with high universality.

[0041] The operation of server 12 in calculating user role preference features based on user historical data can be implemented based on an artificial neural network (ANN) model. For example, it can include, but is not limited to, any one of the following: deep neural networks, recurrent neural networks, recursive neural networks, convolutional neural networks, graph convolutional networks (GCN), and deep generative models, or models derived from one or more of the above deep learning models. This embodiment does not impose any limitations.

[0042] Next, server 12 can use the calculated answer encoding features and the role preference features to calculate the personalized answer corresponding to the user's question.

[0043] When calculating personalized answers to user questions using answer encoding features and role preference features, a regression model can be used to calculate the vocabulary distribution probability under the influence of these various features, and the specific vocabulary included in the answer can be determined based on the vocabulary distribution probability. This part will be described in detail in subsequent embodiments and will not be elaborated here.

[0044] Figure 1 The diagram illustrates the personalized answers generated by server 12 for different users, such as... Figure 1As shown, based on historical data analysis of User 1, when the user's preference for mobile phone camera performance is found to be high, answers related to mobile phone camera performance can be generated when the user asks performance-related questions. Similarly, based on historical data analysis of User 2, when the user's preference for mobile phone gaming performance is found to be high, answers related to the mobile phone's gaming performance can be generated when the user asks performance-related questions.

[0045] In this embodiment, when the server performs automatic question-and-answer calculations for user questions initiated by an object, it can calculate answer encoding features based on the user question and the object's description information, and calculate the user's role preference features based on the user question and the user's historical data. Furthermore, it can comprehensively consider data from multiple dimensions to perform automatic question-and-answer calculations, thereby matching answers that meet the user's personalized needs. Based on this implementation method, the object's description information and the user's historical data can be fully utilized, improving the effectiveness of the automatic question-and-answer calculation results and helping to meet the user's actual question-and-answer needs.

[0046] In some exemplary embodiments, server 12 may generate personalized answers based on a Personalized Answer Generation (PAGE) model.

[0047] like Figure 2 As shown, the PAGE model includes:

[0048] The Basic Encoder-decoder Architecture (BEDA) module is used to generate generic, non-personalized answers.

[0049] The Persona History Incorporation (PHI) module is used to incorporate users' historical data as external knowledge for knowledge-level user preference modeling.

[0050] The Persona Preference Modeling (PPM) module is used to capture aspect-level user preferences using topic models and dynamically provide a persona lexicon based on user preferences.

[0051] The Persona Information Summarizer (PIS) module is used to summarize user preference information from multiple perspectives to generate personalized answers, and to generate personalized answers based on user preference information from multiple perspectives. (hereinafter referred to as PIS)

[0052] Next, we will combine Figure 2The PAGE model shown here further illustrates the question-and-answer calculation method provided in the embodiments of this application.

[0053] In this embodiment, predicting the answer to a user's question can be based on one round of prediction operations or multiple rounds of prediction operations. In some cases, the answer contains a single word, in which case one round of prediction operations is sufficient to obtain the answer. In other cases, the answer may contain multiple words, requiring multiple rounds of prediction operations. Each prediction round predicts one word, and the words predicted in multiple rounds are concatenated to obtain the complete answer. The word prediction algorithms for each round have certain similarities; the following explanation will use the prediction operation of any round (i.e., round t) as an example.

[0054] Optionally, when calculating the answer encoding features corresponding to the user question based on the user question and the object's description information, the server 12 can calculate the answer word features for the current round based on the association features between the user question and the description information fragment, as well as the user question and existing answer information. Specifically, if the current prediction round is not the first prediction round, the existing answer may include: the answer predicted in the previous round, i.e., the answer output in round t-1; if the current prediction round is the first prediction round, the existing answer may include: the answer referencing information for the current round. That is, if the existing answer is empty, referencing information (such as the referencing symbol of the answer) can be used as the existing answer to facilitate subsequent calculations.

[0055] When calculating the answer encoding features, server 12 can calculate the association features between the user question and the description information of the object as a factual expression of the description information, and can also calculate the association features between the user question and existing answers as the question encoding features of the current prediction round.

[0056] Next, server 12 can determine the answer coding features for the current round based on the correlation features between the factual expression of the descriptive information and the question coding features of the current prediction round.

[0057] The operation of calculating the factual representation of descriptive information described in the above embodiments can be implemented based on the self-attention encoder layer and the bi-attention encoder layer in BEDA.

[0058] Optionally, server 12 can input the user question and description information fragment into BEDA, and obtain the self-attention features of the user question and the attention features of the description information fragment in the self-attention encoder layer. The self-attention features can be calculated based on a multi-head attention algorithm. This calculation process can be expressed by the following formula:

[0059] C*=MultiHead(X*,X*,X*) (1)

[0060] O*=LayerNorm(FFN(C*)+X*) (2)

[0061] Where *∈{q, f1, f2, ..., fk}, LayerNorm() represents the computation of the normalization layer, FFN() represents the computation of the feedforward network, MultiHead() represents the multi-head attention computation, C* represents intermediate computation results, and O* represents the output of the self-attention encoder layer.

[0062] Where, when * = q, X q The text representing the user's input question, O q This represents the self-attention feature of the user question output by the self-attention encoder layer.

[0063] Where * = fk, X fk The text representing the descriptive information fragment of the input, O fk This represents the self-attention feature of the descriptive information fragment fk output by the self-attention encoder layer.

[0064] Next, in the bidirectional attention encoder layer, the bidirectional attention features of the self-attention features of the user question and the self-attention features of the descriptive information can be calculated as a factual representation of the descriptive information. This factual representation of the descriptive information is used to express the deep correlation between the user question and the descriptive information fragment.

[0065] Taking any descriptive information fragment in the descriptive information as an example, when calculating the bidirectional attention features, the similarity between the self-attention features of the user question and the self-attention features of the descriptive information fragment can be calculated first. Based on the similarity, the attention weight of the user question and the attention weight of the descriptive information fragment can be determined.

[0066] This similarity can be expressed using a similarity matrix. Taking the k-th descriptive information fragment fk as an example, the process of calculating the similarity between the self-attention features of the user's question and the self-attention features of the descriptive information fragment can be illustrated by the following formula:

[0067]

[0068] Where ω U ∈R 3mh×1 ω U It is a trainable weight vector, o represents element-wise multiplication, [;] represents concatenation, m represents the dimension of the word vector, and R represents the matrix dimension. U fk ∈R Lq×LfWhere Lq represents the sentence length of the user question q, and Lf represents the sentence length of the descriptive information fragment f.

[0069] The method for calculating attention weights based on the similarity matrix can be referenced in the following formula:

[0070] A fk =softmax(U fk (4-1)

[0071]

[0072] Among them, A fk and B fk It is a row and column normalized similarity vector, A fk B represents the attention weight of the user question on the descriptive information fragment fk. fk To describe the attention weight of information fragment fk to the user's question, softmax() represents the regression calculation.

[0073] Based on the attention weights calculated above, the user question and the descriptive information fragment can be weighted to obtain the weighted user question and descriptive information fragment, as shown in the following formula:

[0074]

[0075]

[0076] in, This indicates the weighted user issue; fk represents the weighted descriptive information fragment.

[0077] Next, the weighted user questions and descriptive information fragments are concatenated to obtain the factual expression of the descriptive information, as shown in the following formula:

[0078]

[0079] BiAttention() represents bidirectional attention computation.

[0080] The operation of calculating the problem coding features of the current round described in the above embodiments can be implemented based on the decoder layer in BEDA.

[0081] Optionally, at the decoder layer, the self-attention features of existing answers can be calculated, and the multi-head attention features between the self-attention features of the user question and the self-attention features of the existing answers can be calculated as the question encoding features for the current prediction round. As shown in the following formula:

[0082] M a=MultiHead(X a X a X a (7)

[0083] M q =MultiHead(M a O q O q (8)

[0084] Among them, X a M represents the text where the answer is already provided. a M represents the self-attention feature of an existing answer. q The multi-head attention features represent the relationship between the self-attention features of the user's question and the self-attention features of the existing answers, i.e., the question encoding features of the current prediction round.

[0085] Next, server 12 can express Y based on the factual information described in this description. f The problem encoding feature M of the current prediction round q The correlation features between them determine the answer encoding features for the current round.

[0086] Optionally, server 12 may, at the decoder layer, determine the answer encoding feature M for the current round based on the association features between the factual representation of the descriptive information and the question encoding features of the current prediction round. f :

[0087] M f =MultiHead(M q Y f Y f (9)

[0088] Odec1=FFN(M f (10)

[0089] Among them, M f Let M represent the answer encoding feature, and Odec1 represent the answer encoding feature M output by the decoder. f The decoding result. Where Y f The concatenated result Y of the factual expressions representing the descriptive information fragments involved in the calculation. f =[Y f1 Y f2 , ..., Y fk ], Y f ∈R kLf×mh .

[0090] Obtain the decoder's encoding feature M of the answer for the current round. fAfter decoding the result Odec1, server 12 can map Odec1 to the global vocabulary to obtain the vocabulary distribution probability P. v (a t The global vocabulary refers to a general vocabulary provided for different users, unaffected by user preferences. Optionally, in this embodiment, the operation of mapping the decoding results of the answer encoding features to the global vocabulary can be implemented based on a pre-trained regression model, as shown in the following formula:

[0091] P v (a t = softmax(W) v Odec1+b v (11)

[0092] Among them, W v b v W represents the learnable mapping parameters during the training phase. v ∈R mh×|V| and b v ∈R |V| |V| represents the vocabulary size of this global vocabulary. Where a t Used to refer to the predicted answer in the t-th prediction round.

[0093] In some exemplary embodiments, the user's role preference characteristics, calculated based on the user's historical data, may include at least one of the user's knowledge-level preference characteristics, aspect-level preference characteristics, and vocabulary-level preference characteristics. The optional calculation methods for these various role preference characteristics will be described below.

[0094] Among them, knowledge-level preference features are features obtained by extracting users' personal cognitive experiences from their historical data as prior knowledge.

[0095] Optionally, after obtaining a large number of historical data fragments from the user, the server 12 can search the large number of historical data fragments based on the user's question to retrieve historical data fragments associated with the user's question, and obtain historical data for calculating personalized answers: H = {h1, h2, ..., hk}.

[0096] In the PAGE model, when merging historical data fragments of a user, PHI can utilize BEDA to calculate the correlation features between the user's question and the historical data, thereby obtaining the user's knowledge-level preference feature Y. h .

[0097] Optionally, the association feature between the user question and the historical data segment can be implemented as a bidirectional attention feature between the self-attention feature of the user question and the self-attention feature of the historical data segment, as shown in the following formula:

[0098] C hk =MultiHead(X hk X hk X hk (12)

[0099] O hk =LayerNorm(FFN(C hk )+X hk (13)

[0100] Y hk =BiAttention(O hk O q (14)

[0101] Among them, X hk This represents the text of the input historical data segment hk, C hk Y represents the intermediate calculation process, and Ohk represents the self-attention feature of the historical data segment hk. hk This represents the bidirectional attention feature between the user question and the historical data segment hk. After calculating the bidirectional attention features between the user question and the historical data segments h1, h2, ..., hk respectively, and concatenating the features obtained from each historical data segment, the user's knowledge-level preference feature Y can be obtained. h ={Y1, Y2...Y hk Y h ∈R kLf ×mh .

[0102] In this embodiment, aspect-level preference features are used to characterize a user's preference for one or more aspects of an object. For example, they can characterize a user's preference for a camera lens, a camera brand, or a camera weight.

[0103] In this embodiment, lexical preference features are used to characterize a user's preferences in word choice, i.e., the user's language style.

[0104] exist Figure 2In the PAGE model shown, server 12 can calculate the user's aspect-level preference features for the object based on the distribution characteristics of the vocabulary contained in the historical data across at least one descriptive aspect of the object, using the PPM. Server 12 can also calculate the user's lexical-level preference features based on the distribution characteristics of the vocabulary contained in the historical data within the global lexicon, using the PPM. Based on the lexical-level preference features, the user's role lexicon can be determined.

[0105] PPM can employ a neural topic model to capture users' latent aspect-level preference features and dynamically provide a role vocabulary based on user preferences. The following will provide an illustrative explanation of the topic modeling process.

[0106] Server 12 can input a set of historical data fragments H = {h1, h2, ..., hk} into PPM. PPM can process the historical data fragments into a bag-of-words (BoW) distribution, which is represented as d.

[0107] like Figure 2 As shown, in the variable reasoning process based on neural networks, the input to the neural network is the bag-of-words distribution d. In the neural network, the bag-of-words distribution d is first encoded into continuous Gaussian variables μ and σ using a multilayer perceptron (MLP), as shown in the following formula:

[0108] μ=FFNμ(FFNd(d)) (15-1)

[0109] log(σ 2 )=FFNσ(FFNd(d)) (15-2)

[0110] Next, construct a system with mean μ and σ 2 The Gaussian prior distribution x = N(μ, σ) is the diagonal of the covariance matrix. 2 ):

[0111] x~N(μ,σ 2 (16)

[0112] Based on the Gaussian prior distribution, an unbiased gradient estimator can be constructed for the Gaussian prior distribution x using the re-parameterization trick.

[0113] Next, label a latent variable θ, θ∈R K K represents the number of aspect categories, a latent variable derived from a Gaussian prior distribution. A softmax-based regression operation is performed on the prior distribution x to compute the probability distribution of user preference categories at the aspect level (i.e., the aspect distribution):

[0114] θ = softmax(W) x x+b x (17)

[0115] Among them, W x and b x The weights and bias parameters are for a learnable linear transformation. During training, samples obtained from Gaussian random sampling can be used to reparameterize x = μ + ∈ · σ, where ∈ ∈ N(0, I), and (θ|d) is used to represent the variable inference process that approximates the true posterior (θ|d). For a certain word w in the bag of roles... n The aspect assignment (i.e., the aspect category to which a word belongs) is represented as z. n The process of reconstructing the bag-of-words distribution d using a neural network to parameterize the aspect distribution is as follows:

[0116] z n ~Multi(θ), n∈[1, |V bow |] (18)

[0117]

[0118] Where, β zn V represents the aspect distribution of the vocabulary given aspect assignment zn. bow This refers to the size of the BoW bag-of-words vocabulary. Next, the marginal probability is calculated using the following formula:

[0119]

[0120] To explicitly model user preferences, the aspect distribution of word β can be constructed by multiplying the aspect vector T and the Bow word vector V, where T∈R. K×mh V∈R |Vbow|×mh As shown in the following formula:

[0121] β = softmax(TV) T ) (twenty one)

[0122] Here, the aspect vector T is a feature vector used to represent aspect-based features. β∈R K×|Vbow| This can be viewed as a semantic similarity matrix between words in the category and role vocabulary. Next, based on the similarity matrix, aspect-level user preferences and vocabulary-level user preferences can be separated.

[0123] Next, the reconstructed user-personalized bag-of-words distribution d′ can be determined, where d′ = θ·β. Based on the user-personalized bag-of-words distribution, the user's role vocabulary can be determined.

[0124] By combining the aspect distribution θ with the aspect vector T, an aspect-level user preference vector Y can be constructed for each user. t (i.e., aspect-level preference features), as shown in the following formula:

[0125] Y t =θ·T (22-1)

[0126] Combining the bag-of-words distribution d′ with the word vectors V from BoW allows for the construction of a vocabulary-level user preference vector Y for each user. p (i.e., lexical-level preference features), as shown in the following formula:

[0127] Y p =d′·V (22-2)

[0128] In this process, user aspect-level preference features are further integrated into the answer decoding stage to model aspect-level user preferences; lexical preference vectors are used to enhance the generated answers by incorporating user-preferred language styles.

[0129] During the training phase, PPM can use an objective function L based on the definition of a lower bound for negative variables. PPM The topic model based on neural networks is optimized as shown in the following formula:

[0130]

[0131] Among them, D KL This indicates the calculation of KL divergence (Kullback-Leibler, also known as relative entropy divergence), and P() indicates the probability calculation.

[0132] Wherein, given a sample The latent variable zn can be integrated as follows:

[0133]

[0134] The above objective function L PPM It can be represented as:

[0135]

[0136] Based on the descriptions of the above embodiments, for each round of prediction operation, after obtaining the answer encoding features and role preference features of this round, the personalized answer corresponding to the user's question can be calculated using the answer encoding features and role preference features. The following will provide an exemplary description.

[0137] Optionally, the answer encoding features of the current round can be input into a regression model to obtain the vocabulary distribution probability, as described in Formula 11 above. In the following embodiments, for ease of description and distinction, the vocabulary distribution probability generated based on the answer encoding features illustrated in Formula 11 is described as the first vocabulary distribution probability.

[0138] Next, server 12 can further fuse the answer encoding features of the current prediction round, the user's knowledge-level preference features, and the user's aspect-level preference features based on BEDA in the PAGE model to obtain fused features.

[0139] Optionally, the calculation of the fusion feature can be based on a multi-head attention algorithm. BEDA can calculate the multi-head attention feature between the answer encoding feature of the current prediction round and the user's knowledge-level preference feature to obtain the answer encoding feature influenced by knowledge-level preference; and calculate the multi-head attention feature between the answer encoding feature influenced by knowledge-level preference and the user's aspect-level preference feature to obtain the fusion feature. As shown in the following formula:

[0140] M h =MultiHead(M f Y h Y h (26)

[0141] M p =MultiHead(M h Y t Y t (27)

[0142] O dec 2 = FFN(M p (28)

[0143] Among them, M f Y is the encoding feature of the answer for the current prediction round calculated in the preceding steps. h Y represents the user's knowledge-level preference characteristics. t M represents the aspect-level preference features of users. h M represents the answer encoding features influenced by knowledge-level preferences. p Used to represent intermediate calculation processes, O dec 2 represents the calculated fusion feature.

[0144] Next, optionally, server 12 can calculate the vocabulary distribution probability influenced by role preferences based on the PIS in the PAGE model, according to the fusion feature and the expanded vocabulary, which is described here as the second vocabulary distribution probability. In calculating the second vocabulary distribution probability, the user's role preference features are comprehensively considered, which can map the user's preferences to the answer vocabulary generation process, so that the final generated answer meets the user's personalized needs.

[0145] After obtaining the distribution probabilities of the first and second word lists, the comprehensive word list distribution probability of the current round can be determined based on the distribution probabilities of the first and second word lists. Based on the comprehensive word list distribution probability, words whose distribution probabilities meet the set conditions can be identified as the words predicted in the current round.

[0146] The following will provide an exemplary description of an optional implementation method for calculating the distribution probability of the second vocabulary.

[0147] Optionally, in PIS, the fused features can be input into a regression model to obtain the original words in the existing word sources as the attention weights of the answer words in the current round; the existing word sources include: the user's question, the historical data, the descriptive information and / or the role word list corresponding to the user's word-level preference features.

[0148] In other words, the attention weights can be calculated for the original words in the user's question as the answer words in the current round; the attention weights can be calculated for each word in the historical data segment as the answer words in the current round; the attention weights can be calculated for each word in the descriptive information segment as the answer words in the current round; and the attention weights can be calculated for each preferred word in the user's role vocabulary as the answer words in the current round. As shown in the following formula:

[0149]

[0150] Where W*, W*s, ω*, and b* are the parameters that need to be learned, and s t This represents the fused feature of the decoder output in round t, i.e., the O of the output in round t of prediction. dec 2; This represents the calculated attention weights. Where, when * = q, Y... * =Oq, to ​​avoid symbol confusion.

[0151] Where * = q, the fusion feature s can be... t And the characteristics Y of this user problem q (i.e. O) q Input the regression model and calculate the first attention weight for each word in the user's question as the answer word in the current round, i.e.:

[0152] Where * = h, the fusion feature s can be... t And the user's knowledge-level preference feature Y h Input the regression model and calculate the second weight of each word in the historical data as the answer word in the current round, i.e.

[0153] Where * = f, the fusion feature s can be... t And the factual expression Y of the descriptive information fragment of the object. f Input the regression model and calculate the third attention weight for each word in the descriptive information as the answer word in the current round, i.e.

[0154] Where * = p, the fusion feature s can be... t And the user's lexical-level preference features Y p Input the regression model and calculate the fourth attention weight for each preferred word in the user's role vocabulary, which serves as the answer word for the current round.

[0155] Next, based on the attention weights of the original words in the existing vocabulary source as answer words in the current round and the probability distribution of the same words in the expanded vocabulary list, the attention-based probability distribution of the original words in the existing vocabulary source can be calculated. The expanded vocabulary list consists of words from the global vocabulary list, words from the user's question, words from the object's description information, words from the user's historical data, and words from the role vocabulary list. As shown in the following formula:

[0156]

[0157]

[0158]

[0159]

[0160] Among them, P q (a t ) represents the attention-based lexical probability distribution of words in the user question, P f (a t ) represents the attention-based word probability distribution of words in the descriptive information, P h (a t ) represents the attention-based word probability distribution of words in historical data, P p (a t ) represents the attention-based word probability distribution of the vocabulary in the character vocabulary. Let P q (a tTaking the calculation process of ) as an example, Let represent the attention weight of the i-th word in the user question, which corresponds to the ω-th word in the extended vocabulary. The attention weight of the i-th word in the user question is superimposed with the probability distribution of the i-th word in the extended vocabulary to obtain the attention-based probability distribution of the i-th word in the user question.

[0161] Based on the calculations of the above embodiments, the probability distribution of the first vocabulary and the multi-probability distribution matrix corresponding to the probability distribution of the second vocabulary can be determined, as shown in the following formula:

[0162] P all (a t )=[P v (a t ), P q (a t ), P f (a t ), P h (a t ), P p (a t (33)

[0163] It is worth noting that since the method provided in this application embodiment determines the comprehensive vocabulary distribution probability based on the probabilities calculated from multiple dimensions, the weight information of the probabilities calculated from multiple dimensions can be calculated before calculating the comprehensive vocabulary distribution probability.

[0164] Optionally, server 12 can first calculate the context vector of the original words in the existing word source as the answer words in the current round, based on the attention weights of the original words in the existing word source. That is, based on the attention weights... Calculate the context vector As shown in the following formula:

[0165]

[0166] When * = q, the characteristic Y of the user problem can be calculated. q The transpose of the matrix and the first attention weight The product yields the context vector for each word in the problem.

[0167] When * = h, the user's knowledge-level preference feature Y is calculated. h The transpose matrix and the second attention weight The product of these terms yields the context vector corresponding to each word in the historical data.

[0168] When * = f, the factual representation Y of the descriptive information fragment of the object is calculated. f The transpose of the matrix and the third attention weight The product of these terms yields the context vector corresponding to each word in the descriptive information.

[0169] When * = p, the user's lexical-level preference feature Y is calculated. p The transpose of the matrix and the fourth attention weight The product of these terms yields the context vector corresponding to each of the user's preferred words, i.e.

[0170] Next, the fused feature and the context vector corresponding to each existing word source can be input into the regression model to obtain multi-probability weight information, as shown in the following formula:

[0171]

[0172] Among them, W γ and b γ γ is a multi-view pointer scalar used to determine the weights of each view in the probability distribution, and is a parameter that needs to be learned during the training phase.

[0173] Based on the multi-probability distribution matrix and multi-probability weight information, server 12 can calculate the comprehensive vocabulary distribution probability for the current round. Optionally, server 12 can calculate the product of the transpose of the multi-probability distribution matrix and the multi-probability weight information to obtain the comprehensive vocabulary distribution probability for the current round, as shown in the following formula:

[0174] P(a t ) = P all (a t ) T γ (35)

[0175] After obtaining the distribution probability of the comprehensive vocabulary, words whose distribution probability meets the set conditions are selected as the answer words for the current round. The set conditions for distribution probability to meet the conditions may include: words with high distribution probability rankings or distribution probabilities greater than a set threshold; this embodiment does not impose any restrictions.

[0176] It is also worth noting that in some optional embodiments, server 12 may employ an end-to-end approach to jointly learn the potential role preference information and the personalized answer generation process. In addition to the objective function illustrated in Equation 25 above, the objective function required for the end-to-end training process may also include: an objective function for training the answer generation process, as shown in the following formula:

[0177]

[0178] To ensure the diversity of outcome preference embeddings in PPM, a regularization term L is added to the objective function L in PPM. R To distinguish the aspect vector T for each aspect preference:

[0179] L R =||T·T T -I|| (37)

[0180] Where I is the identity matrix. Based on the above, the overall objective function for end-to-end model training can be expressed as L PPM L QA and L R The weighted sum is shown in the following formula:

[0181] L = L PPM +λ1(L QA +λ1L R (38)

[0182] Where λ1 and λ2 are hyperparameters used to balance the loss.

[0183] The question-answering computing system provided in the above and following embodiments of this application can be applied to various application scenarios. For example, pre-sales consultation scenarios, after-sales service scenarios, and medication consultation scenarios, etc., and this application does not limit it in this regard.

[0184] The following example illustrates the pre-sales consultation scenario.

[0185] In a pre-sales consultation scenario, before making a purchase, a user can ask questions about the product through terminal device 11. Server 12 can answer the user's questions. In some embodiments, the e-commerce platform can provide users with pre-sales consultation services via robot customer service through terminal device 11. The product page displayed on terminal device 11 may include a customer service consultation entry point. When a user triggers this entry point, they can enter the robot customer service page and ask a question. After receiving the user's question, the robot customer service can send it to server 12. Server 12 can calculate the answer encoding features corresponding to the user's question based on the user's question and the product description information, and calculate the user's role preference features based on the user's historical data. Based on the answer encoding features and role preference features, server 12 can calculate a personalized answer that better meets the user's consultation intent.

[0186] For example, when a user views product information on a clothing product page provided by terminal device 11, they can access the chatbot customer service page through the customer service inquiry portal provided on that clothing product page and ask the chatbot customer service representative, "How is this garment?" The chatbot customer service representative can then forward the user's question to server 12.

[0187] When server 12 receives the user's question, it can calculate a personalized partial answer based on the PAGE model. When calculating role preference features, if the user's historical data contains a large amount of comments and questions about clothing styles, the role preference features calculated by server 12 may include clothing style preferences. If the user's historical comments and questions contain a lot of colloquial expressions, the role preference features calculated by server 12 may include colloquial language style preferences. Based on the answer encoding features and role preference features, the personalized answer calculated by server 12 for the user's question could be: "Waist-cinching style, a slimming wonder!"

[0188] If a user has a large number of comments and questions about clothing fabrics in their historical data, and the user prefers a more formal writing style, then based on the answer coding features and role preference features, the personalized answer calculated by server 12 for the user's question could be: This garment is made of pure cotton fabric, which is soft to the touch and breathable.

[0189] After calculating the personalized answer, the server 12 can send the personalized answer to the terminal device 11, which will then display it in the form of a dialogue on the robot customer service page.

[0190] In other embodiments, the e-commerce platform can provide users with an "Ask Everyone" consultation service through terminal device 11. Based on this consultation service, users can ask questions to other users who have purchased the product, and the e-commerce platform can also answer the user's questions. Typically, the product page displayed on terminal device 11 may include a question entry point. After triggering this entry point, the user can enter the question they want to ask and submit it. After obtaining the user's question, terminal device 11 can send the user's question to server 12. On the one hand, server 12 can send the question message to other users who have purchased the product to obtain answers from other users regarding the question. On the other hand, server 12 can calculate personalized answers for users based on the PAGE model described in the aforementioned embodiments.

[0191] For example, when a user views product information on a fruit product page provided by terminal device 11, they can access a question page through the "Ask Everyone" control provided on the product page and enter the question: "Is the quality of this fruit product good?". Terminal device 11 can then send the user's question to server 12.

[0192] When server 12 receives the user's question, it can send the question to the terminal devices of other users who have previously purchased the fruit product and retrieve their answers. After retrieving the answers, server 12 can send them to terminal device 11. Terminal device 11 can then display the other users' answers in the answer display area corresponding to the user's question on the "Ask Everyone" page.

[0193] Simultaneously, server 12 can calculate a personalized answer to the user's question based on the PAGE model. When calculating the user's role preference characteristics, if the user's historical data contains a large amount of comments and questions about the freshness of fruit products, the role preference characteristics calculated by server 12 may include a preference for freshness. If the user's historical comments and questions contain a large amount of colloquial expressions, the role preference characteristics calculated by server 12 may include a preference for colloquial language style. Based on answer encoding characteristics and role preference characteristics, the personalized answer calculated by server 12 for the user's question could be: "It's so fresh! It's like it was just picked from the tree!" If the user's historical data contains a large amount of comments and questions about the taste of fruit products, and the user's language style prefers more formal written expressions, then based on answer encoding characteristics and role preference characteristics, the personalized answer calculated by server 12 for the user's question could be: "This fruit is slightly sour at first taste, but has a sweet aftertaste."

[0194] After calculating the personalized answer, server 12 can send the personalized answer to terminal device 11. Terminal device 11 can then display the personalized answer as one of the answers to the user's question on the "Ask Everyone" page.

[0195] The above embodiments describe the system structure and system functions of the question-answering calculation system provided in this application. The following sections will provide a detailed description of the question-answering calculation method provided in the embodiments of this application in conjunction with the accompanying drawings.

[0196] Figure 3 This is a flowchart illustrating a question-answering calculation method provided in an exemplary embodiment of this application. This embodiment can be based on... Figure 1 The question-answering computation system implementation shown is primarily described from the server's perspective. For example... Figure 3 As shown, the method includes:

[0197] Step 301: Receive user questions sent by the terminal device, wherein the user questions are initiated by the user for a specific object.

[0198] Step 302: Calculate the answer encoding features corresponding to the user question based on the user question and the description information of the object.

[0199] Step 303: Calculate the user's role preference characteristics based on the user's historical data.

[0200] Step 304: Calculate the personalized answer corresponding to the user's question using the answer encoding features and the role preference features.

[0201] In some exemplary embodiments, one method of calculating the answer encoding features corresponding to the user question based on the user question and the description information of the object may include: calculating the association features between the user question and the description information as a factual expression of the description information; and,

[0202] Calculate the association features between the user's question and existing answers as the question coding features for the current prediction round; wherein, the existing answers include: the answer predicted in the previous round or the answer referencing information for the current prediction round; determine the answer coding features for the current round based on the association features between the factual expression of the descriptive information and the question coding features for the current prediction round.

[0203] In some exemplary embodiments, calculating the association features between the user question and the description information as a way of expressing the facts of the description information may include: for any fragment of description information in the description information, obtaining the self-attention features of the user question and the attention features of the description information fragment; calculating the bidirectional attention features of the self-attention features of the user question and the attention features of the description information fragment as a way of expressing the facts of the description information fragment.

[0204] In some exemplary embodiments, when predicting the answer for the current round, one way to calculate the association features between the user question and the existing answers as the answer encoding features for the current prediction round may include: obtaining the self-attention features of the user question and the self-attention features of the existing answers; calculating the multi-head attention features between the self-attention features of the user question and the self-attention features of the existing answers as the answer encoding features for the current prediction round.

[0205] In some exemplary embodiments, one way to calculate the user's role preference characteristics based on the user's historical data may include: calculating the user's knowledge-level preference characteristics based on the correlation characteristics between the user's questions and the historical data; and / or, calculating the user's aspect-level preference characteristics for the object based on the distribution characteristics of the vocabulary contained in the historical data on at least one descriptive aspect of the object; and / or, calculating the user's role vocabulary and vocabulary-level preference characteristics based on the distribution characteristics of the vocabulary contained in the historical data in a global vocabulary.

[0206] In some exemplary embodiments, one way to calculate the user's knowledge-level preference features based on the association features between the user's question and the historical data may include: obtaining the self-attention features of any historical data segment in the historical data; calculating the bidirectional attention features between the self-attention features of the historical data segment and the self-attention features of the user's question to obtain the user's knowledge-level preference features.

[0207] In some exemplary embodiments, one way to calculate the user's aspect-level preference features for the object based on the distribution characteristics of the vocabulary contained in the historical data on at least one descriptive aspect of the object may include: obtaining the vocabulary contained in the historical data as the user's role bag; inputting the role bag into a topic model to obtain the distribution characteristics of each word in the role bag on at least one descriptive aspect of the object; and determining the user's aspect-level preference features based on the distribution characteristics of each word in the role bag on at least one descriptive aspect of the object.

[0208] In some exemplary embodiments, one method for calculating personalized answers to user questions using the answer encoding features and the role preference features may include: inputting the answer encoding features of the current round into a regression model to obtain a first vocabulary distribution probability; fusing the answer encoding features of the current prediction round, the user's knowledge-level preference features, and the user's aspect-level preference features to obtain a fused feature; calculating a second vocabulary distribution probability based on the fused feature and an expanded vocabulary; determining a comprehensive vocabulary distribution probability for the current round based on the first and second vocabulary distribution probabilities; and determining words whose distribution probabilities satisfy a set condition based on the comprehensive vocabulary distribution probability, as the words predicted for the current round.

[0209] In some exemplary embodiments, one way to fuse the answer encoding features of the current prediction round, the user's knowledge-level preference features, and the user's aspect-level preference features to obtain fused features may include: calculating a multi-head attention feature between the answer encoding features of the current prediction round and the user's knowledge-level preference features to obtain answer encoding features influenced by knowledge-level preferences; and calculating a multi-head attention feature between the answer encoding features influenced by knowledge-level preferences and the user's aspect-level preference features to obtain the fused features influenced by both knowledge-level preferences and aspect-level preferences.

[0210] In some exemplary embodiments, one way to calculate the probability distribution of the second vocabulary list based on the fusion features and the expanded vocabulary list may include: inputting the fusion features into a regression model to obtain the attention weights of the original words in the existing vocabulary sources as answer words in the current round; the existing vocabulary sources include: the user question, the historical data, the description information and / or the role vocabulary list corresponding to the user's vocabulary-level preference features; and calculating the attention-based probability distribution of the original words in the existing vocabulary sources based on the attention weights of the original words in the existing vocabulary sources as answer words in the current round and the probability distribution of the same words in the expanded vocabulary list.

[0211] In some exemplary embodiments, the fused features are input into a regression model to obtain the attention weight of each word in the existing word sources as the answer word in the current round, including at least one of the following: inputting the fused features and the features of the user question into the regression model to calculate the first attention weight of each word in the user question as the answer word in the current round; inputting the fused features and the user's knowledge-level preference features into the regression model to calculate the second attention weight of each word in the historical data as the answer word in the current round; inputting the fused features and the factual expression of the descriptive information into the regression model to calculate the third attention weight of each word in the descriptive information as the answer word in the current round; and inputting the fused features and the user's word-level preference features into the regression model to calculate the fourth attention weight of each preferred word of the user as the answer word in the current round.

[0212] In some exemplary embodiments, one way to determine the comprehensive vocabulary distribution probability of the current round based on the first vocabulary distribution probability and the second vocabulary distribution probability may include: determining the multi-probability distribution matrix corresponding to the first vocabulary distribution probability and the second vocabulary distribution probability; calculating the context vector of the original words in the existing vocabulary sources as the answer words of the current round based on the attention weights of the original words in the existing vocabulary sources as the answer words of the current round; inputting the fused features and the context vector of the original words in the existing vocabulary sources as the answer words of the current round into a regression model to obtain multi-probability weight information; and obtaining the comprehensive vocabulary distribution probability of the current round based on the multi-probability distribution matrix and the multi-probability weight information.

[0213] In some exemplary embodiments, calculating the context vector of the original words in the existing word source as the answer words in the current round, based on the attention weights of the original words in the existing word source as the answer words in the current round, includes at least one of the following: calculating the transpose matrix of the features of the user's question and multiplying it by the first attention weight to obtain the context vector corresponding to each word in the question; calculating the transpose matrix of the user's knowledge-level preference features and multiplying it by the second attention weight to obtain the context vector corresponding to each word in the historical data; calculating the transpose matrix of the factual expression of the descriptive information fragment and multiplying it by the third attention weight to obtain the context vector corresponding to each word in the descriptive information; and calculating the transpose matrix of the user's word-level preference features and multiplying it by the fourth attention weight to obtain the context vector corresponding to each preferred word of the user.

[0214] In this embodiment, when performing automatic question-and-answer calculations on user questions initiated by an object, answer encoding features can be calculated based on the user question and the object's description information. Furthermore, user role preference features can be calculated based on the user question and the user's historical data. This allows for comprehensive consideration of data from multiple dimensions to perform automatic question-and-answer calculations, thereby matching answers that meet the user's personalized needs. Based on this implementation method, the object's description information and the user's historical data can be fully utilized, improving the effectiveness of automatic question-and-answer calculation results and better meeting the user's actual question-and-answer needs.

[0215] Figure 4 This is a flowchart of a question-answering calculation method provided by another exemplary embodiment of this application. This embodiment can be based on... Figure 1 The question-answering computation system implementation shown is primarily described from the server's perspective. For example... Figure 4 As shown, the method includes:

[0216] Step 401: Display the first interface, which includes at least one question entry point for the object.

[0217] Step 402: Respond to the user's question entry trigger operation and display a question and answer interface adapted to the triggered question entry.

[0218] Step 403: Obtain the user question initiated by the user regarding the object through the question and answer interface, and send the user question to the server so that the server can determine a personalized answer to the user question based on the user question, the description information of the object, and the user's historical data.

[0219] In this embodiment, the user's question and the object's description information can be used to calculate the answer encoding features, while the user's question and the user's historical data can be used to calculate the user's role preference features. Furthermore, automatic question-answering calculations can be performed by comprehensively considering data from multiple dimensions to match answers that meet the user's personalized needs, thereby improving the effectiveness of the automatic question-answering calculation results.

[0220] The question-and-answer calculation method provided in this embodiment can be implemented by the terminal device in the question-and-answer calculation system described in the above embodiments. For technical details and beneficial effects not described in detail in this embodiment, please refer to the content related to the terminal device described in the above embodiments, which will not be repeated here.

[0221] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 301 to 304 can be device A; or the execution subject of steps 301 and 302 can be device A, and the execution subject of steps 303 and 304 can be device B; and so on.

[0222] It should be noted that some processes described in the above embodiments and accompanying drawings include multiple operations appearing in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear in this document, or they may be executed in parallel. The operation numbers, such as 401, 402, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should also be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0223] The above describes optional embodiments of the question-answering calculation method applicable to server 12, such as... Figure 5 As shown, in some exemplary embodiments, the server 12 for performing the above-described question-and-answer calculation method may include: a memory 501, a processor 502, a graph communication component 503, and a power supply component 504.

[0224] Memory 501 can be configured to store various other data to support operations on server 12. Examples of this data include instructions for any application or method operating on server 12, contact data, phonebook data, messages, pictures, videos, etc. Memory can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0225] In this embodiment, the memory 501 is used to store one or more computer instructions.

[0226] Processor 502, coupled to memory 501, is configured to execute one or more computer instructions in memory 501 for: receiving user questions sent by terminal devices via graph communication component 503, the user questions being initiated by a user for an object; calculating answer encoding features corresponding to the user questions based on the user questions and description information of the objects; calculating the user's role preference features based on the user's historical data; and calculating personalized answers corresponding to the user questions using the answer encoding features and the role preference features.

[0227] In some exemplary embodiments, when the processor 502 calculates the answer encoding features corresponding to the user question based on the user question and the description information of the object, it is specifically used to: calculate the association features between the user question and the description information as a factual expression of the description information; and calculate the association features between the user question and existing answers as question encoding features for the current prediction round; wherein, the existing answers include: the answer predicted in the previous round of the current round or the answer referencing information of the current prediction round; and determine the answer encoding features for the current round based on the association features between the factual expression of the description information and the question encoding features for the current prediction round.

[0228] In some exemplary embodiments, when the processor 502 calculates the association features between the user question and the description information as a factual representation of the description information, it is specifically configured to: for any fragment of description information in the description information, obtain the self-attention features of the user question and the attention features of the description information fragment; calculate the bidirectional attention features of the self-attention features of the user question and the attention features of the description information fragment as a factual representation of the description information fragment.

[0229] In some exemplary embodiments, when the processor 502 is predicting the answer for the current round, it calculates the association features between the user question and the existing answers as the question encoding features for the current prediction round. Specifically, this is done by: obtaining the self-attention features of the user question and the self-attention features of the existing answers; and calculating the multi-head attention features between the self-attention features of the user question and the self-attention features of the existing answers as the question encoding features for the current prediction round.

[0230] In some exemplary embodiments, when the processor 502 calculates the user's role preference features based on the user's historical data, it is specifically configured to: calculate the user's knowledge-level preference features based on the correlation features between the user's questions and the historical data; and / or, calculate the user's aspect-level preference features for the object based on the distribution features of the vocabulary contained in the historical data on at least one descriptive aspect of the object; and / or, calculate the user's role vocabulary and vocabulary-level preference features based on the distribution features of the vocabulary contained in the historical data in the global vocabulary.

[0231] In some exemplary embodiments, when the processor 502 calculates the user's knowledge-level preference features based on the association features between the user's question and the historical data, it specifically performs the following: for any historical data segment in the historical data, obtain the self-attention feature of the historical data segment; calculate the bidirectional attention feature between the self-attention feature of the historical data segment and the self-attention feature of the user's question, and obtain the user's knowledge-level preference features.

[0232] In some exemplary embodiments, when the processor 502 calculates the user's aspect-level preference features for the object based on the distribution characteristics of the words contained in the historical data on at least one descriptive aspect of the object, it specifically performs the following steps: obtaining the words contained in the historical data as the user's role bag of words; inputting the role bag of words into a topic model to obtain the distribution characteristics of each word in the role bag of words on at least one descriptive aspect of the object; and determining the user's aspect-level preference features based on the distribution characteristics of each word in the role bag of words on at least one descriptive aspect of the object.

[0233] In some exemplary embodiments, when the processor 502 calculates the personalized answer corresponding to the user's question using the answer encoding features and the role preference features, it specifically performs the following steps: inputting the answer encoding features of the current round into a regression model to obtain a first vocabulary distribution probability; fusing the answer encoding features of the current prediction round, the user's knowledge-level preference features, and the user's aspect-level preference features to obtain a fused feature; calculating a second vocabulary distribution probability based on the fused feature and an expanded vocabulary; determining a comprehensive vocabulary distribution probability for the current round based on the first vocabulary distribution probability and the second vocabulary distribution probability; and determining words whose distribution probabilities satisfy a set condition based on the comprehensive vocabulary distribution probability, as the words predicted for the current round.

[0234] In some exemplary embodiments, when the processor 502 fuses the answer encoding features of the current prediction round, the user's knowledge-level preference features, and the user's aspect-level preference features to obtain fused features, it specifically performs the following: calculates the multi-head attention features between the answer encoding features of the current prediction round and the user's knowledge-level preference features to obtain answer encoding features influenced by knowledge-level preferences; calculates the multi-head attention features between the answer encoding features influenced by knowledge-level preferences and the user's aspect-level preference features to obtain the fused features influenced by both knowledge-level preferences and aspect-level preferences.

[0235] In some exemplary embodiments, when the processor 502 calculates the probability distribution of the second vocabulary based on the fusion features and the expanded vocabulary, it specifically performs the following steps: inputting the fusion features into a regression model to obtain the attention weights of the original words in the existing vocabulary sources as answer words in the current round; the existing vocabulary sources include: the user question, the historical data, the description information, and / or the role vocabulary corresponding to the user's vocabulary-level preference features; and calculating the attention-based probability distribution of the original words in the existing vocabulary sources based on the attention weights of the original words in the existing vocabulary sources as answer words in the current round and the probability distribution of the same words in the expanded vocabulary.

[0236] In some exemplary embodiments, when the processor 502 inputs the fused features into a regression model to obtain the attention weight of each word in the existing word source as the answer word in the current round, it specifically performs at least one of the following operations: inputting the fused features and the features of the user question into the regression model to calculate the first attention weight of each word in the user question as the answer word in the current round; inputting the fused features and the user's knowledge-level preference features into the regression model to calculate the second attention weight of each word in the historical data as the answer word in the current round; inputting the fused features and the factual expression of the descriptive information into the regression model to calculate the third attention weight of each word in the descriptive information as the answer word in the current round; and inputting the fused features and the user's word-level preference features into the regression model to calculate the fourth attention weight of each of the user's preferred words as the answer word in the current round.

[0237] In some exemplary embodiments, when the processor 502 determines the comprehensive vocabulary distribution probability of the current round based on the first vocabulary distribution probability and the second vocabulary distribution probability, it specifically performs the following steps: determining the multi-probability distribution matrix corresponding to the first vocabulary distribution probability and the second vocabulary distribution probability; calculating the context vector of the original words in the existing vocabulary sources as the answer words of the current round based on the attention weights of the original words in the existing vocabulary sources as the answer words of the current round; inputting the fused features and the context vector of the original words in the existing vocabulary sources as the answer words of the current round into a regression model to obtain multi-probability weight information; and obtaining the comprehensive vocabulary distribution probability of the current round based on the multi-probability distribution matrix and the multi-probability weight information.

[0238] In some exemplary embodiments, when the processor 502 calculates the context vector of the original words in the existing word source as the answer words in the current round based on the attention weights of the original words in the existing word source as the answer words in the current round, it specifically performs at least one of the following operations: calculating the transpose matrix of the features of the user's question and multiplying it by the first attention weight to obtain the context vector corresponding to each word in the question; calculating the transpose matrix of the user's knowledge-level preference features and multiplying it by the second attention weight to obtain the context vector corresponding to each word in the historical data; calculating the transpose matrix of the factual expression of the descriptive information fragment and multiplying it by the third attention weight to obtain the context vector corresponding to each word in the descriptive information; and calculating the transpose matrix of the user's word-level preference features and multiplying it by the fourth attention weight to obtain the context vector corresponding to each preferred word of the user.

[0239] In this embodiment, when performing automatic question-and-answer calculations on user questions initiated by an object, answer encoding features can be calculated based on the user question and the object's description information. Furthermore, user role preference features can be calculated based on the user question and the user's historical data. This allows for comprehensive consideration of data from multiple dimensions to perform automatic question-and-answer calculations, thereby matching answers that meet the user's personalized needs. Based on this implementation method, the object's description information and the user's historical data can be fully utilized, improving the effectiveness of automatic question-and-answer calculation results and better meeting the user's actual question-and-answer needs.

[0240] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed, can implement the steps in the method embodiments described above that can be executed by a server.

[0241] like Figure 6 As shown, in some exemplary embodiments, the terminal device 11 for performing the above-described question-and-answer calculation method may include: a memory 601, a processor 602, a communication component 603, a display component 604, an audio component 605, and a power supply component 606.

[0242] Memory 601 can be configured to store various other data to support operation on terminal device 11. Examples of such data include instructions for any application or method operating on terminal device 11, contact data, phonebook data, messages, pictures, videos, etc. Memory can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0243] In this embodiment, the memory 601 is used to store one or more computer instructions.

[0244] Processor 602, coupled to memory 601, is configured to execute one or more computer instructions in memory 601 for: displaying a first interface via display component 604, the first interface including at least one question entry for an object; responding to a user-initiated question entry trigger operation, displaying a question-and-answer interface adapted to the triggered question entry; obtaining a user question initiated by the user for the object through the question-and-answer interface, and sending the user question to a server so that the server can determine a personalized answer to the user question based on the user question, the object's description information, and the user's historical data.

[0245] The display component 604 includes a liquid crystal display (LCD) and a touch panel (TP). If the display component 604 includes a touch panel, it can be implemented as a touchscreen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0246] In an optional embodiment, audio component 605 is stored as output and / or input audio signals. For example, audio component 605 includes a microphone (MIC) configured to receive external audio signals when the device containing audio component 605 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 601 or transmitted via communication component 603. For example, in some optional embodiments, terminal device 11 may receive user-input voice signals through the microphone of audio component 605, perform voice recognition and semantic recognition on the voice signals to determine user questions or user behavior characteristics related to user questions. In some embodiments, audio component 605 also includes a speaker for outputting audio signals. For example, it may be used to output answers corresponding to user questions.

[0247] In an alternative embodiment, power supply component 606 provides power to various components of terminal device 11. The power supply component may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to terminal device 11.

[0248] In this embodiment, the user's question and the object's description information can be used to calculate the answer encoding features, while the user's question and the user's historical data can be used to calculate the user's role preference features. Furthermore, automatic question-answering calculations can be performed by comprehensively considering data from multiple dimensions to match answers that meet the user's personalized needs, thereby improving the effectiveness of the automatic question-answering calculation results.

[0249] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed, can implement the steps in the method embodiments described above that can be executed by a terminal device.

[0250] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0251] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0252] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0253] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0254] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0255] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0256] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0257] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0258] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A question-and-answer calculation method, characterized in that, include: Receive user questions sent by terminal devices, wherein the user questions are initiated by users for a specific object; Calculate the association features between the user question and the description information of the object, and use them as a factual representation of the description information; Furthermore, the association features between the user's question and existing answers are calculated as the question encoding features for the current prediction round; wherein, the existing answers include: the answer predicted in the previous round or the answer referencing information for the current prediction round; and the answer encoding features for the current round are determined based on the association features between the factual expression of the descriptive information and the question encoding features for the current prediction round. Based on the user's historical data, the user's role preference features are calculated; the role preference features include at least one of the following: knowledge-level preference features, aspect-level preference features representing the user's preference for at least one aspect of an object, and lexical-level preference features representing the user's preference in terms of word usage; the knowledge-level preference features are features obtained by extracting the user's personal cognitive experience from the user's historical data as prior knowledge. The answer encoding features of the current round are input into the regression model to obtain the first vocabulary distribution probability; the answer encoding features of the current prediction round, the user's knowledge-level preference features, and the user's aspect-level preference features are fused to obtain fused features; based on the fused features and the expanded vocabulary, the second vocabulary distribution probability is calculated; based on the first vocabulary distribution probability and the second vocabulary distribution probability, the comprehensive vocabulary distribution probability of the current round is determined; based on the comprehensive vocabulary distribution probability, the words whose distribution probabilities meet the set conditions are determined as the words predicted in the current round.

2. The method according to claim 1, characterized in that, Calculating the association features between the user question and the description information, as a factual representation of the description information, includes: For any fragment of the description information, obtain the self-attention feature of the user question and the attention feature of the description information fragment; Calculate the bidirectional attention features of the self-attention features of the user question and the attention features of the descriptive information fragment, and use them as the factual representation of the descriptive information fragment.

3. The method according to claim 1, characterized in that, When predicting the answer for the current round, the association features between the user's question and existing answers are calculated and used as the question encoding features for the current prediction round, including: Obtain the self-attention features of the user's question and the self-attention features of the existing answers; Calculate the multi-head attention feature between the self-attention feature of the user's question and the self-attention feature of the existing answer, and use it as the question encoding feature for the current prediction round.

4. The method according to claim 1, characterized in that, Based on the user's historical data, calculate the user's role preference characteristics, including: Calculate the user's knowledge-level preference characteristics based on the correlation features between the user's questions and the historical data; and / or, Based on the distribution characteristics of the vocabulary contained in the historical data across at least one descriptive aspect of the object, calculate the user's aspect-level preference features for the object; and / or, Based on the distribution characteristics of the words contained in the historical data in the global lexicon, the user's role lexicon and vocabulary-level preference characteristics are calculated.

5. The method according to claim 4, characterized in that, Based on the correlation characteristics between the user's questions and the historical data, the user's knowledge-level preference characteristics are calculated, including: For any historical data segment in the historical data, obtain the self-attention feature of the historical data segment; Calculate the bidirectional attention features between the self-attention features of the historical data segment and the self-attention features of the user's question to obtain the user's knowledge-level preference features.

6. The method according to claim 4, characterized in that, Based on the distribution characteristics of the vocabulary contained in the historical data across at least one descriptive aspect of the object, calculate the user's aspect-level preference features for the object, including: Obtain the vocabulary contained in the historical data and use it as the user's role vocabulary bag; Input the bag-of-words for the character into the topic model to obtain the distribution characteristics of each word in the bag-of-words for the character on at least one descriptive aspect of the object; Based on the distribution characteristics of each word in the bag of words for the role across at least one descriptive aspect of the object, the aspect-level preference characteristics of the user are determined.

7. The method according to claim 1, characterized in that, The answer encoding features of the current prediction round, the user's knowledge-level preference features, and the user's aspect-level preference features are fused to obtain fused features, including: Calculate the multi-head attention feature between the answer encoding feature of the current prediction round and the user's knowledge-level preference feature to obtain the answer encoding feature influenced by knowledge-level preference; Calculate the multi-head attention features between the answer encoding features influenced by knowledge-level preferences and the user's aspect-level preference features to obtain the fusion features influenced by both knowledge-level preferences and aspect-level preferences.

8. The method according to claim 2, characterized in that, Based on the fusion features and the expanded vocabulary, the distribution probability of the second vocabulary is calculated, including: The fused features are input into the regression model to obtain the original words in the existing word sources as the attention weights of the answer words in the current round; the existing word sources include: the user question, the historical data, the descriptive information and / or the role word list corresponding to the user's word-level preference features; Based on the attention weights of the original words in the existing word sources as the answer words in the current round and the probability distribution of the same words in the expanded word list, calculate the attention-based probability distribution of the original words in the existing word sources.

9. The method according to claim 8, characterized in that, The fused features are input into a regression model to obtain the attention weight of each word in the existing word source as the answer word in the current round, including at least one of the following: The fused features and the features of the user question are input into the regression model to calculate the first attention weight of each word in the user question as the answer word in the current round. The fusion features and the user's knowledge-level preference features are input into the regression model to calculate each word in the historical data as the second attention weight of the answer word in the current round. The fused features and the factual representation of the descriptive information are input into the regression model, and each word in the descriptive information is calculated as the third attention weight of the answer word in the current round. The fusion features and the user's lexical-level preference features are input into the regression model to calculate the fourth attention weight of each preferred word of the user as the answer word in the current round.

10. The method according to claim 9, characterized in that, Based on the distribution probabilities of the first and second vocabularies, the comprehensive vocabulary distribution probability for the current round is determined, including: Determine the multi-probability distribution matrix corresponding to the first vocabulary distribution probability and the second vocabulary distribution probability; Based on the attention weights of the original words in the existing word sources as the answer words in the current round, calculate the context vectors of the original words in the existing word sources as the answer words in the current round; The fused features and the original words in the existing word sources are used as the context vectors of the answer words in the current round and input into the regression model to obtain multi-probability weight information. Based on the multi-probability distribution matrix and the multi-probability weight information, the comprehensive vocabulary distribution probability of the current round is obtained.

11. The method according to claim 10, characterized in that, Based on the attention weights of the original words in the existing word source as the answer words in the current round, the context vectors of the original words in the existing word source as the answer words in the current round are calculated, including at least one of the following: Calculate the transpose matrix of the features of the user question and multiply it by the first attention weight to obtain the context vector corresponding to each word in the question; Calculate the product of the transpose of the user's knowledge-level preference features and the second attention weight to obtain the context vector corresponding to each word in the historical data; Calculate the product of the transpose matrix of the factual representation of the descriptive information fragment and the third attention weight to obtain the context vector corresponding to each word in the descriptive information; The transpose of the user's lexical-level preference features is multiplied by the fourth attention weight to obtain the context vector corresponding to each preferred word of the user.

12. A question-and-answer calculation method, characterized in that, include: Display a first interface, which includes at least one question entry point for the object; Responding to user-initiated questions, the system displays a Q&A interface adapted to the triggered question entry. The question-and-answer interface is used to obtain user questions raised by users regarding the object, and the user questions are sent to the server so that the server can calculate the correlation features between the user questions and the description information of the object, as a factual expression of the description information; Furthermore, the association features between the user's question and existing answers are calculated as the question encoding features for the current prediction round; wherein, the existing answers include: the answer predicted in the previous round or the answer referencing information for the current prediction round; the answer encoding features for the current round are determined based on the association features between the factual expression of the descriptive information and the question encoding features for the current prediction round; and the user's role preference features are calculated based on the user's historical data; the role preference features include at least one of the following: knowledge-level preference features, aspect-level preference features representing the user's preference for at least one aspect of an object, and lexical-level preference features representing the user's preference in terms of word usage; knowledge-level The preference features are obtained by extracting the user's personal cognitive experience from the user's historical data as prior knowledge; the answer encoding features of the current round are input into the regression model to obtain the first vocabulary distribution probability; the answer encoding features of the current prediction round, the user's knowledge-level preference features, and the user's aspect-level preference features are fused to obtain the fused features; the second vocabulary distribution probability is calculated based on the fused features and the expanded vocabulary; the comprehensive vocabulary distribution probability of the current round is determined based on the first and second vocabulary distribution probabilities; and the words whose distribution probabilities satisfy the set conditions are determined based on the comprehensive vocabulary distribution probability and used as the words predicted in the current round.

13. A server, characterized in that, include: Memory, processor, and communication components; The memory is used to store one or more computer instructions; The processor is configured to execute one or more computer instructions for performing the steps of the method according to any one of claims 1-11.

14. A terminal device, characterized in that, include: Memory, processor, and communication components; The memory is used to store one or more computer instructions; The processor is configured to execute one or more computer instructions for performing the steps of the method of claim 12.

15. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed, it can perform the steps of the method according to any one of claims 1-12.

Images