Text argumentation strategy analysis method, analysis device and computer readable storage medium
By performing argument component sequence encoding and dynamic matrix graph convolution operations on the argument text, combined with a variational inference model, the problems of lack of diversity and low accuracy in argument strategy analysis are solved, and more efficient argument strategy prediction and response analysis are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH SHENZHEN GRADUATE SCHOOL
- Filing Date
- 2022-09-06
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies lack the ability to characterize the diversity of argumentation strategies and have a single sequence prediction pattern in the analysis of argumentation strategies, resulting in low analysis accuracy.
The target argument text is encoded into an argument component sequence encoder. By obtaining the probability distribution of historical argument components and the vector of the target historical argument components, and combining dynamic matrix and graph convolution operations, the probability distribution of the argument strategy of the current argument component is predicted. Finally, the response strategy is predicted using a variational inference argument strategy sequence prediction model.
It improves the accuracy of argumentation strategy analysis, better preserves the overall information of the target argumentation text, and enhances the accuracy of the diverse characterization and prediction of argumentation strategies.
Smart Images
Figure CN116306582B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of natural language processing, and in particular relates to a text argumentation strategy analysis method, analysis device and computer-readable storage medium. Background Technology
[0002] Current analysis of argumentation strategies generally simplifies it into a classification task, categorizing simple argumentation strategies employed in sentences. However, this approach fails to adequately capture the complex and diverse nature of argumentation strategies. Furthermore, argumentation prediction is merely sequence-to-sequence, exhibiting a singular prediction pattern and lacking diversity. Summary of the Invention
[0003] This application provides a text argumentation strategy analysis method, analysis device, and computer-readable storage medium, which can improve the accuracy of argumentation strategy analysis of target argumentation texts.
[0004] The first aspect of this application provides a method for analyzing textual argumentation strategies. The method includes: acquiring a target argumentation text; using an argumentation component sequence encoder to encode the target argumentation text into a sequence of argumentation components, the sequence of argumentation components including multiple argumentation components, and the multiple argumentation components being ordered in chronological order; for the current argumentation component, acquiring the probability distribution corresponding to each historical argumentation component in the sequence of argumentation components, wherein the probability distribution corresponding to the historical argumentation components includes the probability of each argumentation strategy corresponding to the historical argumentation component; determining the historical prediction vector corresponding to the current argumentation component based on the probability distribution corresponding to each historical argumentation component; acquiring the target vector corresponding to the target historical argumentation component, wherein the target historical argumentation component is adjacent to the current argumentation component; determining the target vector corresponding to the current argumentation component based on the target vector corresponding to the target historical argumentation component, the historical prediction vector, and the representation vector of the current argumentation component; and predicting the target vector corresponding to the current argumentation component to obtain a probability distribution for predicting the current argumentation component.
[0005] A second aspect of this application provides an analysis apparatus, which includes a processor, a memory, and a communication circuit. The processor is coupled to the memory and the communication circuit, respectively. The memory stores program data, and the processor executes the program data in the memory to implement the steps in the above method.
[0006] A third aspect of this application provides a computer-readable storage medium storing a computer program that can be executed by a processor to implement the steps in the above-described method.
[0007] The beneficial effect is that this application uses an encoder to encode the argument text into a sequence of argument components, and uses the predicted probability distribution of the argument strategy of the argument component to help predict the probability distribution of the argument strategy of the next argument component, which can improve the accuracy of the argument strategy analysis of the target argument text. Attached Figure Description
[0008] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:
[0009] Figure 1 This is a flowchart illustrating one implementation method of the textual argumentation strategy analysis method of this application;
[0010] Figure 2 This is a schematic diagram of the user interface of the tool for analyzing and predicting the argumentation strategy of this application;
[0011] Figure 3 This is a schematic diagram of the argumentation strategy analysis model based on attention dynamic graph convolutional neural network in this application;
[0012] Figure 4 yes Figure 1 A flowchart illustrating step S140;
[0013] Figure 5 yes Figure 1 A flowchart illustrating step S160;
[0014] Figure 6 This is a schematic diagram of the structure of the argument strategy sequence prediction model based on variational reasoning in this application;
[0015] Figure 7 yes Figure 6 Enlarged view of point A in the middle;
[0016] Figure 8 This is a partial flowchart illustrating one implementation method of the textual argumentation strategy analysis method of this application;
[0017] Figure 9 This is a schematic diagram of one embodiment of the analysis device of this application;
[0018] Figure 10 This is a schematic diagram of one embodiment of the computer-readable storage medium of this application. Detailed Implementation
[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0020] It should be noted that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.
[0021] See Figure 1 In one embodiment of this application, the textual argumentation strategy analysis method includes:
[0022] S110: Obtain the target argument text.
[0023] Specifically, the textual argumentation strategy analysis method of this application is executed by an analysis device, which has an argumentation strategy analysis and prediction tool, and its user interface is as follows: Figure 2 As shown, the user first inputs a target argument text in area A. Then, the analysis device analyzes the argument strategies used in the target argument text and displays the analysis results in area B. At the same time, it also predicts the argument strategies to respond to the target argument text and displays the prediction results in area C.
[0024] S120: Using an argument component sequence encoder, the target argument text is encoded into an argument component sequence, which includes multiple argument components, and the multiple argument components are ordered in a forward-to-back order.
[0025] Specifically, in combination Figure 3 The argument component sequence encoder can be a recurrent neural network, a hierarchical attention network, or a pre-trained language model, etc.
[0026] Among them, an argumentative component is the smallest text segment in the target argumentative text that has argumentative significance. The way to combine the encoded argumentative components of the target argumentative text is called an argumentative component sequence, such as...Figure 3 As shown, the sequence of argument components is represented as (c1, c2, ..., c...). m The coded argument components are arranged in order from front to back, where c i This represents the vector representation of the i-th argument component.
[0027] S130: For the current argument component, obtain the probability distribution corresponding to each historical argument component in the argument component sequence, wherein the probability distribution corresponding to each historical argument component includes the probability of each argument strategy corresponding to the historical argument component.
[0028] Specifically, one of the objectives of this application is to determine the probability distribution corresponding to each argument component. This is achieved by determining the probability distribution of each argument component sequentially from beginning to end. That is, the probability distribution corresponding to the first argument component is determined first, then the second, then the third, and so on. In other words, all argument components are traversed in a sequential manner to obtain the probability distribution corresponding to each argument component. The argument component currently being traversed is defined as the current argument component, and previously traversed argument components are defined as historical argument components of the current argument component.
[0029] This application summarizes common argumentation strategies to better reflect real-world situations, identifying five common argumentation strategy types and their sequence patterns, as shown in Table 1 below:
[0030] Table 1 Typical Argumentation Strategy Types and Sequence Patterns
[0031]
[0032]
[0033] Among these, acknowledging the opponent's viewpoint means partially agreeing with it, laying the groundwork for further refutation; identifying contradictions means finding flaws or contradictions in the opponent's viewpoint, analyzing and evaluating these contradictions to reduce the credibility of their argument; reductio ad absurdum means deriving a false conclusion from the refuted topic or evidence. It involves assuming the opponent's argument is true and then deriving a false conclusion, thus concluding that the argument is invalid; proposing a new viewpoint means presenting and elaborating on one's own new viewpoint to make one's position more credible; and advancing the argument means strengthening or progressing the existing argument to create a more generalized or widespread argument.
[0034] The probability distribution for each argument component includes the probability of the argument strategy used by the argument component as a pre-set argument strategy. The pre-set argument strategies are not limited to the five argument strategies mentioned above. For example, the five argument strategies can be further subdivided into three argument strategies each, resulting in a probability distribution with 15 probability values.
[0035] S140: Determine the historical prediction vector corresponding to the current argument component based on the probability distribution corresponding to each historical argument component.
[0036] Specifically, the obtained historical prediction vectors can carry information about the probability distributions corresponding to each historical argument component. Since the argumentation strategy used by the current argument component is influenced by the identified strategies, combining the historical prediction vectors to obtain the probability distribution for the current argument component's predictions can improve accuracy.
[0037] See Figure 4 In this embodiment, step S140 specifically includes:
[0038] S141: Determine the initial historical vector based on the probability distribution corresponding to each historical argument component.
[0039] In this implementation, the initial historical vector is determined according to the following formula:
[0040]
[0041] Where p is the initial historical vector. Let j be the j-th value in the probability distribution corresponding to the i-th historical argument component. Let C be the vector representation of the argument strategy corresponding to the j-th value in the probability distribution corresponding to the i-th historical argument component, where C is the number of argument strategies and k is the number of historical argument components.
[0042] S142: For each historical argument component, obtain the attention score corresponding to the historical argument component based on the vector of the historical argument component and the initial historical vector.
[0043] In this embodiment, for the historical argumentation component i, its corresponding attention score α is determined according to the following formula. i :
[0044] f(c i ,p)=tanh(c i ·W p ·p T +b p )
[0045]
[0046] Among them, c i Let W be the vector representation of historical argument component i, n be the total number of argument components, and W be the vector representation of the argument component i. p b p All of these are pre-set model parameters.
[0047] S143: Determine the dynamic matrix based on the attention score corresponding to each historical argument component.
[0048] Specifically, in this embodiment, the dynamic matrix is determined according to the following formula:
[0049]
[0050] Among them, A i,j Let α be the element in the i-th row and j-th column of the dynamic matrix A, where i is an integer between 1 and k, and j is an integer between 1 and k, where k equals the number of historical argument components, and α i Let be the attention score corresponding to the i-th historical argument component.
[0051] S144: Perform graph convolution operation using the dynamic matrix and the historical prediction vector corresponding to the target historical argument component to obtain the historical prediction vector corresponding to the current argument component.
[0052] Specifically, graph convolution operations are used to obtain the historical prediction vector corresponding to the current argument component. The specific process is expressed by the following formula:
[0053]
[0054]
[0055] in, It is the symmetric adjacency matrix of the dynamic matrix A normalized to g. l-1 The historical prediction vector w corresponding to the target historical argument component (the previous argument component of the current argument component, which has been analyzed). l Here are the learnable neural network parameters, and ReLU is the activation function, where... The meaning is to put matrix g l Add all column vectors together to obtain the historical prediction vector corresponding to the current debate component. When calculating the historical prediction vector corresponding to the first argument component, g in the formula... l-1 It is the original vectorized representation of the first argument component.
[0056] S150: Obtain the target vector corresponding to the target historical argument component, wherein the target historical argument component is adjacent to the current argument component.
[0057] Specifically, the target historical argument component is the argument component preceding the current argument component, and the target vector corresponding to the current argument component is represented as follows: The target vector corresponding to the target historical argument component is represented as follows:
[0058] S160: Determine the target vector corresponding to the current argument component based on the target vector corresponding to the target historical argument component, the historical prediction vector, and the representation vector of the current argument component.
[0059] Specifically, the target vector corresponding to the current argument component is represented as: The target vector corresponding to the target historical argument component is represented as follows: The historical prediction vector corresponding to the current argument component is represented as follows: The representation vector of the current argument component is c. i .
[0060] In this embodiment, refer to Figure 5 Step S160 specifically includes:
[0061] S161: Concatenate the target vector corresponding to the target historical argument component with the representation vector of the current argument component to obtain the first vector.
[0062] Specifically, the formula for calculating the first vector is: The process of concatenating two vectors is existing technology and will not be elaborated here.
[0063] S162: Concatenate the target vector corresponding to the target historical argument component with the historical prediction vector to obtain the second vector.
[0064] Specifically, the formula for calculating the second vector is:
[0065] S163: Merge the first vector and the second vector to obtain the target vector corresponding to the current argument component.
[0066] Specifically, the first vector With the second vector Add a decoder (i.e.) Figure 3 The argument strategy sequence decoder (e.g., a gated recurrent unit (GRU)) obtains the target vector representation corresponding to the current argument component as follows: Of course, other decoders can also be chosen, such as Long Short-Term Memory (LSTM) networks.
[0067] In this embodiment, the decoder can employ a gating mechanism for fusion. and To obtain the target vector corresponding to the current argument component. Specifically, in one application scenario, it can be determined according to the following formula.
[0068]
[0069] Where ReLU is the activation function, W c b c W p and b p These are all learnable neural network parameters, where σ represents the sigmoid function.
[0070] S170: Predict the target vector corresponding to the current argument component to obtain the probability distribution for predicting the current argument component.
[0071] Specifically, after obtaining the target vector corresponding to the current argument component... Then, the probability distribution y corresponding to the current argument component is determined using the following formula:
[0072]
[0073] Where W and b are learnable neural network parameters, and softmax represents the softmax function.
[0074] As can be seen from the above, this application combines the probability distributions of historical argument components when determining the probability distribution of the current argument component, thus better preserving the global information of the target argument text. Furthermore, the construction method of the dynamic matrix is highly flexible, improving the accuracy of the overall analytical approach.
[0075] The above describes the process of obtaining the probability distribution corresponding to the current argument component. After traversing each argument component using the above method, the probability distribution corresponding to each argument component can be obtained. At this point, for each argument component, based on its corresponding probability distribution, its target argument strategy can be determined, thus identifying the argument strategy adopted by the argument component as the target argument strategy. Specifically, in determining the target argument strategy, the maximum probability value is found in the corresponding probability distribution, and the argument strategy corresponding to this maximum probability value is then identified as the target argument strategy.
[0076] After the above steps, each argument component has its corresponding target argument strategy. The following describes the process of predicting the argument strategy used by the text that responds to the target argument text.
[0077] In the subsequent prediction process, it is necessary to use, such as Figure 6 The argument strategy sequence prediction model based on variational reasoning shown below will be combined with... Figure 6 , Figure 7 as well as Figure 8 Introduction to the prediction process:
[0078] S210: For each argument component, obtain the input feature vector corresponding to the argument component based on the encoded feature vector corresponding to the argument component and the encoded feature vector corresponding to the target argument strategy of the argument component.
[0079] Specifically, for argumentation component i, its corresponding encoded feature vector is fused with the encoded feature vector corresponding to its target argumentation strategy to obtain the input feature vector d corresponding to argumentation component i. i The set of input feature vectors corresponding to all argument components can be denoted as d = {d1, d2, ..., dn}. m}, where m is the number of argument components.
[0080] The encoded feature vector corresponding to the argument component can be a vector representation of the argument component, or it can be obtained by encoding the argument component using an encoder. Similarly, the encoded feature vector corresponding to the target argument strategy can be obtained in a similar way.
[0081] S220: Input the input feature vectors corresponding to all argument components into the first encoder to obtain fused features.
[0082] Specifically, {d1,d2,…,d m The first encoder is input to obtain the fused features. The specific processing procedure of the first encoder is not limited in this application.
[0083] S230: For the current argument component, obtain the predicted response strategy vector for the current argument component.
[0084] Specifically, obtain the predicted response strategy vector h of the current argument component t. t-1 .
[0085] Among them, for h t-1 By making predictions, we can obtain the probability distribution of the predicted response strategies for the current argument components. Ultimately, the probability distribution of the predicted response strategy is based on the current argument components. This allows us to obtain the corresponding response strategy for the current argument component, that is, what argument strategy to use to respond to the current argument component.
[0086] S240: Based on the fusion features and the predicted response strategy vector of the current argument component, perform attention mechanism operation to obtain the baseline vector of the current argument component.
[0087] Specifically, the fusion features and the predicted response strategy vector h of the current argument component will be used. t-1 By performing attention mechanism operations, the baseline vector c of the current argument component can be obtained.t .
[0088] S250: The first decoder performs calculations based on the baseline vector and the predicted response strategy vector of the current argument component to obtain the latent variable vector.
[0089] Specifically, the first decoder includes a Prior Network and an Inference Network, where the Inference Network is only needed during the training phase.
[0090] The specific process of step S250 includes: setting the reference vector c t The predicted response strategy vector h of the current argument component t-1 Inputting the Prior Network yields the parameters μ of the normal distribution. p ,logσ p Then, the latent variable vector z of the current argument component is obtained through parameterized probability distribution sampling. t The process can be expressed by the following formula:
[0091]
[0092] τ p ~N(0,1)
[0093]
[0094] in, The meaning is to use h i-1 and c t Input PriorNetwork, get μ p and logσ p .
[0095] S260: The first decoder performs calculations based on the predicted response strategy vector, latent variable vector, and predicted response strategy probability distribution of the current debate component to obtain the predicted response strategy vector of the next debate component. The first decoder makes predictions based on the predicted response strategy vector of the current debate component to obtain the predicted response strategy probability distribution of the current debate component. The predicted response strategy probability distribution of the current debate component includes the probability of each response debate strategy corresponding to the current debate component.
[0096] Specifically, the predicted response strategy vector h of the current argument component... t-1 The latent variable vector z of the current argument component t And the probability distribution of the predicted response strategy for the current argument component. The calculation is performed to obtain the predicted response strategy vector h for the next argument component. t .
[0097] S270: The first decoder predicts the predicted response strategy vector for the next argument component and obtains the probability distribution of the predicted response strategy for the next argument component.
[0098] Specifically, the predicted response strategy vector h for the next argument component. t Make predictions to obtain the probability distribution of the predicted response strategy for the next argument component. That is, the probability value of obtaining the response strategy corresponding to the next argument component is the pre-set response argument strategy.
[0099] Using the above method, we can obtain the probability distribution of the predicted response strategy for each argument component, and then determine the response strategy for each argument component, that is, what response strategy should be used for the corresponding argument component when responding to the target argument text.
[0100] The following is an introduction to... Figure 6 The training process of the variational reasoning-based argument strategy sequence prediction model shown requires the use of an Inference Network during training. Specifically, the training process includes:
[0101] S310: Obtain the sample argument text and the corresponding sample response text.
[0102] Specifically, the sample response text corresponding to the sample argument text is the text that responds to the sample argument text, wherein the sample response text is known.
[0103] S320: The argument component sequence encoder is used to encode the sample argument file into a sample argument component sequence. The sample argument sequence includes multiple sample argument components, and the multiple sample argument components are ordered in order from front to back.
[0104] Specifically, this process is the same as step S120 above, and you can refer to the relevant content above for details, which will not be repeated here.
[0105] S330: For each sample argument component, generate a sample response strategy vector and a sample response strategy probability distribution based on the sample response text.
[0106] Specifically, for each sample argument component, its corresponding sample response strategy vector, sample response strategy probability distribution, and sample response strategy are all known and belong to the labeled data in the training process.
[0107] S340: For each sample argument component, obtain the sample input feature vector corresponding to the sample argument component based on the encoded feature vector corresponding to the sample argument component and the encoded feature vector corresponding to the sample argument strategy of the sample argument component.
[0108] Specifically, this process is the same as step S210 above, and will not be repeated here.
[0109] S350: Input the encoded features corresponding to all sample argument components into the first encoder to obtain sample fusion features.
[0110] Specifically, this process is the same as step S220 above, and will not be repeated here.
[0111] S360: For the current sample argument component, obtain the predicted response strategy vector of the current sample argument component.
[0112] Specifically, obtain the predicted response strategy vector h of the current sample argument component. t-1 .
[0113] S370: The first decoder performs attention mechanism operations based on the fused features and the predicted response strategy vector of the current sample argument component to obtain the sample reference vector of the current sample argument component.
[0114] Specifically, step S370 is similar to step S240, obtaining the sample reference vector c of the current sample argument component. t .
[0115] S380: The first latent variable vector is obtained by using the first decoder to perform calculations based on the sample baseline vector and the predicted response strategy vector of the current sample argument component.
[0116] Specifically, the reference vector c t The predicted response strategy vector h of the current sample argument component t-1 Input the PriorNetwork to obtain the parameters μ of the normal distribution. p ,logσ p Then, the first latent variable vector z of the current sample argument component is obtained through parameterized probability distribution sampling. t The process can be expressed by the following formula:
[0117]
[0118] τ p ~N(0,1)
[0119]
[0120] The specific process of step S380 is the same as that of step S250 above.
[0121] S390: The first decoder performs calculations based on the sample baseline vector, the predicted response strategy vector of the current argument component, and the posterior vector corresponding to the current sample argument component to obtain the second latent variable vector; wherein, the posterior vector corresponding to the current sample argument component is obtained based on the sample response strategy probability distribution of the current sample argument component and the posterior vector corresponding to the next sample argument component.
[0122] Specifically, the sample reference vector c t The predicted response strategy vector h of the current sample argument component t-1 And the posterior vector b corresponding to the current sample argument component. t Inputting into the Inference Network (posterior network) yields the parameters μ of the normal distribution. i ,logσ i Then, the second latent variable vector z′ of the current sample argument component is obtained through parameterized probability distribution sampling. t The process can be expressed by the following formula:
[0123]
[0124] τ i ~N(0,1)
[0125]
[0126] Wherein, the posterior vector b corresponding to the current sample argument component t Yes, based on the probability distribution y of the sample response strategy of the current sample argument component. t (This probability distribution is known) and the posterior vector b corresponding to the next sample argument component. t+1 Obtained.
[0127] S400: Generate divergence values based on the first latent variable vector and the second latent variable vector.
[0128] One of the purposes of training is to make the first hidden variable vector z t and the second hidden variable vector z′ t Getting closer and closer.
[0129] S410: Generate the loss value based on the divergence value.
[0130] Specifically, in this embodiment, the process of generating the loss value is as follows:
[0131]
[0132]
[0133] L KL =-KL(p inference (z′ t |y ＜t ,x)||p prior (z t |y ＜t ,x))
[0134] L=βL generation +(1-β)L KL
[0135] Where L is the final loss value, L KL Let be the divergence value, where β is a pre-set weight value that can be set according to the actual situation and is not restricted here.
[0136] S420: Train the first encoder and the first decoder based on the loss value.
[0137] In this embodiment, incorporating the divergence value into the calculation of the loss value can make the prior probability as close as possible to the posterior probability, thereby improving the accuracy of the model detection.
[0138] See Figure 9 , Figure 9 This is a schematic diagram of one embodiment of the analysis device of this application. The analysis device 200 includes a processor 210, a memory 220, and a communication circuit 230. The processor 210 is coupled to the memory 220 and the communication circuit 230 respectively. The memory 220 stores program data. The processor 210 executes the program data in the memory 220 to implement the steps in any of the above embodiments. The detailed steps can be found in the above embodiments and will not be repeated here.
[0139] The analysis device 200 can be any device with algorithm processing capabilities, such as a computer or mobile phone, and there are no restrictions on it.
[0140] See Figure 10 , Figure 10 This is a schematic diagram of one embodiment of the computer-readable storage medium of this application. The computer-readable storage medium 400 stores a computer program 410, which can be executed by a processor to implement the steps in any of the above methods.
[0141] Specifically, the computer-readable storage medium 400 can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or a device that can store the computer program 410. Alternatively, it can be a server that stores the computer program 410, which can send the stored computer program 410 to other devices for execution, or it can run the stored computer program 410 itself.
[0142] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A method for analyzing textual argumentation strategies, characterized in that, The method includes: Obtain the target argument text; An argument component sequence encoder is used to encode the target argument text into an argument component sequence, the argument component sequence including multiple argument components, and the multiple argument components are ordered in a front-to-back order in the argument component sequence; For the current argument component, obtain the probability distribution corresponding to each historical argument component in the argument component sequence, wherein the probability distribution corresponding to the historical argument components includes the probability of each argument strategy corresponding to the historical argument components; Based on the probability distribution corresponding to each of the historical debate components, determine the historical prediction vector corresponding to the current debate component; Obtain the target vector corresponding to the target historical argument component, wherein the target historical argument component is adjacent to the current argument component; The target vector corresponding to the current argument component is determined based on the target vector corresponding to the target historical argument component, the historical prediction vector, and the representation vector of the current argument component. The target vector corresponding to the current argument component is predicted to obtain the probability distribution for predicting the current argument component.
2. The method according to claim 1, characterized in that, The step of determining the historical prediction vector corresponding to the current argument component based on the probability distribution corresponding to each of the historical argument components includes: The initial historical vector is determined based on the probability distribution corresponding to each of the aforementioned historical argument components; For each historical argument component, the attention score corresponding to the historical argument component is obtained based on the vector of the historical argument component and the initial historical vector; A dynamic matrix is determined based on the attention score corresponding to each of the historical argument components; By performing a graph convolution operation using the dynamic matrix and the historical prediction vector corresponding to the target historical argument component, the historical prediction vector corresponding to the current argument component is obtained.
3. The method according to claim 2, characterized in that, The step of determining the initial historical vector based on the probability distribution corresponding to each of the historical argument components includes: The initial historical vector is determined according to the following formula: in, Let the historical initial vector be... Let j be the j-th value in the probability distribution corresponding to the i-th historical argument component. Let C be the vector representation of the argument strategy corresponding to the j-th value in the probability distribution corresponding to the i-th historical argument component, where C is the number of the argument strategies and k is the number of the historical argument components.
4. The method according to claim 2, characterized in that, The step of determining the dynamic matrix based on the attention score corresponding to each of the historical argument components includes: Construct the dynamic matrix according to the following formula: in, Let be the element in the i-th row and j-th column of the dynamic matrix, where i is an integer between 1 and k, and j is an integer between 1 and k, where k equals the number of historical argument components. Let be the attention score corresponding to the i-th historical argument component.
5. The method according to claim 1, characterized in that, The step of determining the target vector corresponding to the current argument component based on the target vector corresponding to the target historical argument component, the historical prediction vector, and the representation vector of the current argument component includes: The target vector corresponding to the target historical argument component is concatenated with the representation vector of the current argument component to obtain the first vector; The target vector corresponding to the target historical argument component is concatenated with the historical prediction vector to obtain the second vector; The first vector and the second vector are fused together to obtain the target vector corresponding to the current argument component.
6. The method according to claim 1, characterized in that, After predicting the target vector corresponding to the current argument component to obtain the probability distribution for predicting the current argument component, the method further includes: After obtaining the probability distribution corresponding to each of the argument components, the target argument strategy corresponding to each argument component is determined, wherein the probability corresponding to the target argument strategy is the highest in the probability distribution corresponding to the argument component.
7. The method according to claim 6, characterized in that, After obtaining the probability distribution corresponding to each of the argument components and determining the target argument strategy corresponding to each argument component, the method further includes: For each of the argumentation components, the input feature vector corresponding to the argumentation component is obtained based on the encoded feature vector corresponding to the argumentation component and the encoded feature vector corresponding to the target argumentation strategy of the argumentation component. The input feature vectors corresponding to all the argument components are input into the first encoder to obtain the fused features; For the current argument component, obtain the predicted response strategy vector of the current argument component; An attention mechanism is performed based on the fusion features and the predicted response strategy vector of the current argument component to obtain the baseline vector of the current argument component. The first decoder performs calculations based on the baseline vector and the predicted response strategy vector of the current argument component to obtain the latent variable vector; The first decoder performs calculations based on the predicted response strategy vector of the current debate component, the latent variable vector, and the predicted response strategy probability distribution of the current debate component to obtain the predicted response strategy vector of the next debate component. The first decoder makes predictions based on the predicted response strategy vector of the current debate component to obtain the predicted response strategy probability distribution of the current debate component. The predicted response strategy probability distribution of the current debate component includes the probability of each response debate strategy corresponding to the current debate component. The first decoder predicts the predicted response strategy vector of the next argument component to obtain the predicted response strategy probability distribution corresponding to the next argument component.
8. The method according to claim 7, characterized in that, The method further includes: Obtain the sample debate text and the corresponding sample response text; The argument component sequence encoder is used to encode the sample argument text into a sample argument component sequence, the sample argument component sequence including multiple sample argument components, and the multiple sample argument components are ordered in a front-to-back order; For each of the sample debate components, a sample response strategy vector and a sample response strategy probability distribution are generated based on the sample response text. For each of the sample argumentation components, the sample input feature vector corresponding to the sample argumentation component is obtained based on the encoded feature vector corresponding to the sample argumentation component and the encoded feature vector corresponding to the sample argumentation strategy of the sample argumentation component. Input the encoded feature vectors corresponding to all the sample argumentation components into the first encoder to obtain the sample fusion features; For the current sample argument component, obtain the predicted response strategy vector of the current sample argument component; The first decoder performs attention mechanism operations based on the fused features and the predicted response strategy vector of the current sample argument component to obtain the sample reference vector of the current sample argument component; The first latent variable vector is obtained by performing calculations using the first decoder based on the sample baseline vector and the predicted response strategy vector of the current sample argumentation component. The first decoder performs calculations based on the sample baseline vector, the predicted response strategy vector of the current sample debate component, and the posterior vector corresponding to the current sample debate component to obtain the second latent variable vector; wherein, the posterior vector corresponding to the current sample debate component is obtained based on the sample response strategy probability distribution of the current sample debate component and the posterior vector corresponding to the next sample debate component; Generate divergence values based on the first latent variable vector and the second latent variable vector; Based on the divergence value, a loss value is generated; The first encoder and the first decoder are trained based on the loss value.
9. An analytical apparatus, characterized in that, The analysis device includes a processor, a memory, and a communication circuit. The processor is coupled to the memory and the communication circuit. The memory stores program data. The processor executes the program data in the memory to implement the steps of the method as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement the steps of the method as described in any one of claims 1-8.