Methods, apparatuses, and devices related to matching opinion mining

Abstract

The present application discloses a method and apparatus for mining matching viewpoints. Among them, the method includes: obtaining a first text and a second text, where the first text and the second text include sentences; determining a shared sentence vector of the sentences according to the word vectors in the sentences through a shared sentence vector determination module in the matching viewpoint recognition model; judging whether a sentence is an argument according to the shared sentence vector through an argument classification classifier in the matching viewpoint recognition model; and judging whether the sentences of the first text and the sentences of the second text are matching viewpoints according to the shared sentence vector through a matching viewpoint classification classifier in the matching viewpoint recognition model. By adopting this processing method, the multi-task training model combines two subtasks of argument mining and argument pairing to learn a better shared sentence representation, and further mines the one-to-one matching viewpoints in the argumentative text; therefore, the accuracy of mining matching viewpoints can be effectively improved.

Description

Technical Field

[0001] The present application relates to the technical field of argumentation mining, and specifically relates to a method and device for mining matching viewpoints, a method and device for constructing a matching viewpoint recognition model, a debate robot system, and an electronic device. Background Art

[0002] Argumentation Mining is an important task in computational argumentation. Its main goal is to automatically extract arguments from text in order to provide structured data for computational models of argumentation and reasoning engines. Currently, most existing argumentation mining tasks are performed on a single paragraph, such as extracting arguments from legal documents, papers, etc. In addition, existing multi-text based argumentation mining is limited to online discussions or debates on forums, extracting arguments from a large number of articles on online forums.

[0003] In the process of implementing the present invention, the inventor proposed a new task of argument pair mining, which is significantly different from the existing tasks in the field of debate mining. The argument pair (matching viewpoint) mining task can perform argument pair mining on a relatively long debate text (where the arguments in the text have a strong tendency to attack each other). For example, for a paper, given a whole paragraph of review comments from a reviewer (which may include summarizing advantages and disadvantages, asking questions, raising doubts, giving suggestions, etc.), and the response from the paper author to the review comments (which may include answering, refuting, etc.), through argument pair mining, the corresponding argument pairs in the review comments and their responses can be extracted. Argument pair mining is of great significance and can be widely applied to research in the field of argumentation mining, such as application fields like debate robots, e-commerce reviews, and intelligent justice.

[0004] In summary, how to perform argument pair mining on debate texts, and how to further expand applications such as debate robots, e-commerce reviews, and intelligent justice using the results of argument pair mining, have become urgent problems for those skilled in the art to solve. Summary of the Invention

[0005] The present application provides a method for mining matching viewpoints to solve the problem in the prior art that it is impossible to mine the corresponding matching viewpoints in debate texts. The present application also provides a device for mining matching viewpoints, a method and device for constructing a matching viewpoint recognition model, a debate robot system, and an electronic device.

[0006] The present application provides a method for mining matching viewpoints, including:

[0007] Obtaining a first text and a second text including matching viewpoints, where the first text and the second text include sentences;

[0008] Determine the shared sentence vector of a sentence according to the word vectors in the sentence through the shared sentence vector determination module in the view matching recognition model;

[0009] Judge whether a sentence is an argument according to the shared sentence vector through the argument classifier in the view matching recognition model; and, combine the sentences in the first text and the sentences in the second text in pairs, and determine whether the two sentences are matching views according to the shared sentence vector through the matching view classifier in the view matching recognition model.

[0010] Optionally, the shared sentence vector determination module in the view matching recognition model determines the shared sentence vector of a sentence according to the word vectors in the sentence, including:

[0011] Determine the sentence vector according to the word vectors in the sentence and the argument text type of the sentence through the sentence encoder in the shared sentence vector determination module;

[0012] Determine the shared sentence vector according to the sentence vectors of multiple sentences through the sentence vector sharing sub-module in the shared sentence vector determination module.

[0013] Optionally, judge whether a sentence is an argument according to the shared sentence vector through the argument classifier in the view matching recognition model;

[0014] Determine the argument sentences of the first text and the argument sentences of the second text that are paired in pairs according to the argument text type of the sentence and the argument judgment result;

[0015] Judge whether two argument sentences are matching views through the matching view classifier in the view matching recognition model.

[0016] Optionally, it further includes:

[0017] Learn the view matching recognition model from the argument text and the matching view annotation information set.

[0018] Optionally, it further includes:

[0019] Obtain the word vectors through BERT pre-training.

[0020] Optionally, the first text includes: the proponent's argument text, and the second text includes: the opponent's argument text.

[0021] Optionally, the first text includes: the text provided by the plaintiff, and the second text includes: the text provided by the defendant.

[0022] Optionally, the first text includes: the product review text provided by the buyer user, and the second text includes: the comment reply text provided by the seller user.

[0023] The present application also provides a method for constructing a matching opinion recognition model, including:

[0024] Constructing the network structure of the matching opinion recognition model, where the model includes a shared sentence vector determination module, an argumentation classifier, and a matching opinion classifier; the shared sentence vector determination module is used to determine the shared sentence vector of a sentence according to the word vectors in the sentence; the argumentation classifier is used to judge whether a sentence is an argument according to the shared sentence vector; the matching opinion classifier is used to judge whether two sentences are matching opinions according to the shared sentence vector;

[0025] Learning the parameters of the matching opinion recognition model from the argumentative text and the matching opinion annotation information set.

[0026] The present application also provides a method for mining matching opinions, including:

[0027] Obtaining a first text and a second text including matching opinions, where the first text and the second text include sentences;

[0028] Judging whether a sentence is an argument according to the word vectors in the sentence through an argumentation recognition model;

[0029] Combining the argument sentences in the first text and the argument sentences in the second text in pairs, and judging whether two argument sentences are matching opinions according to the word vectors in the sentences through the matching opinion recognition model.

[0030] The present application also provides a device for mining matching opinions, including:

[0031] An argumentative text acquisition unit for obtaining a first text and a second text including matching opinions, where the first text and the second text include sentences;

[0032] A shared sentence vector determination unit for determining the shared sentence vector of a sentence according to the word vectors in the sentence through the shared sentence vector determination module in the matching opinion recognition model;

[0033] A matching opinion determination unit for judging whether a sentence is an argument according to the shared sentence vector through the argumentation classifier in the matching opinion recognition model; and combining the sentences in the first text and the sentences in the second text in pairs, and determining whether two sentences are matching opinions according to the shared sentence vector through the matching opinion classifier in the matching opinion recognition model.

[0034] The present application also provides a debate robot system, including:

[0035] According to the above-mentioned device for mining matching opinions, and a debate device based on matching opinions.

[0036] The present application also provides a computer-readable storage medium, in which instructions are stored. When the instructions run on a computer, the computer is caused to execute the above various methods.

[0037] The present application also provides a computer program product including instructions. When the computer program product runs on a computer, the computer is caused to execute the above various methods.

[0038] Compared with the prior art, the present application has the following advantages:

[0039] The matching view mining method provided by the embodiment of the present application includes obtaining a first text and a second text including matching views; for each sentence in the text, determining a shared sentence vector of the sentence according to the word vectors in the sentence through a shared sentence vector determination module in the matching view recognition model; judging whether the sentence is an argument through an argument classification classifier in the matching view recognition model according to the shared sentence vector; and for each pair of sentences, judging whether the two sentences are matching views according to the shared sentence vector through a matching view classification classifier in the matching view recognition model. This processing method splits the matching view mining task into two subtasks of argument mining and argument pairing, effectively combines the two subtasks through multi-task training of the matching view recognition model, utilizes the complementary content and structural relationship of question and answer in the argumentative text, learns a better shared sentence representation, and then mines the corresponding matching views in the argumentative text; therefore, the accuracy of matching view mining can be effectively improved.

[0040] The method for constructing a matching view recognition model provided by the embodiment of the present application includes constructing a network structure of the matching view recognition model, where the model includes a shared sentence vector determination module, an argument classification classifier, and a matching view classification classifier, and learning the parameters of the matching view recognition model from the argumentative text and the matching view annotation information set; this processing method effectively combines the two subtasks of argument mining and argument pairing through multi-task training of the matching view recognition model, utilizes the complementary content and structural relationship of question and answer in the argumentative text, learns a better shared sentence representation, and then mines the corresponding matching views in the argumentative text; therefore, the accuracy of the model can be effectively improved.

[0041] The matching opinion mining method provided by the embodiment of the present application includes obtaining a first text and a second text including matching opinions; for each sentence in the text, through an argumentation recognition model, judging whether the sentence is an argument according to the word vectors in the sentence; for each pair combination of the argument sentences in the first text and the argument sentences in the second text, through a matching opinion recognition model, judging whether the two argument sentences are matching opinions according to the word vectors in the sentence; this processing method splits the matching opinion mining task into two subtasks: argumentation mining and argumentation pairing, trains two models independently, combines these two models through a pipeline structure, and utilizes the complementary content and structural relationship of the question-and-answer in the argumentative text to mine the corresponding matching opinions in the argumentative text; therefore, the effect of matching opinion mining can be effectively improved.

[0042] The debate robot system provided by the embodiment of the present application includes a matching opinion mining device and a debate device based on matching opinions. For a large number of historical argumentative texts, the matching opinion mining device determines the shared sentence vector of the sentence according to the word vectors in the argumentative text sentence through the shared sentence vector determination module in the matching opinion recognition model; judges whether the sentence is an argument according to the shared sentence vector through the argumentation classifier in the matching opinion recognition model; and combines the argument sentences and the interrogation sentences in pairs, and determines whether the two sentences are matching opinions according to the shared sentence vector through the matching opinion classifier in the matching opinion recognition model; then, through the debate device based on matching opinions, the matching opinion mining result is used for the subsequent speeches of the pros and cons; this processing method splits the matching opinion mining task into two subtasks: argumentation mining and argumentation pairing, and effectively combines these two subtasks through multi-task training of the matching opinion recognition model, utilizes the complementary content and structural relationship of the question-and-answer in the argumentative text, learns a better shared sentence representation, and then mines the corresponding matching opinions in the argumentative text; therefore, the accuracy of matching opinion mining can be effectively improved, and further the accuracy of the automatic speech of the debate robot can be improved. Brief Description of the Drawings

[0043] Figure 1 Flow chart of an embodiment of a matching opinion mining method provided by the present application;

[0044] Figure 2 Text matching opinion diagram of an embodiment of a matching opinion mining method provided by the present application;

[0045] Figure 3 Model structure diagram of an embodiment of a matching opinion mining method provided by the present application;

[0046] Figure 4 Flow chart of an embodiment of a method for constructing a matching opinion recognition model provided by the present application;

[0047] Figure 5 Flow chart of an embodiment of a matching view mining method provided by this application. Detailed implementation manners

[0048] A lot of specific details are set forth in the following description in order to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar promotions without departing from the connotation of this application. Therefore, this application is not limited by the specific implementations disclosed below.

[0049] In this application, a matching view mining method and apparatus, a matching view recognition model construction method and apparatus, a debate robot system, and an electronic device are provided. Various solutions will be described in detail in the following embodiments.

[0050] The first embodiment

[0051] Please refer to Figure 1 , which is a flow chart of an embodiment of the matching view mining method of this application. The execution subject of the method includes but is not limited to a server, and can also be any device capable of implementing the method. In this embodiment, the method may include the following steps:

[0052] Step S101: Obtain a first text and a second text including matching views.

[0053] The first text and the second text constitute a debate text, and the method can perform matching view mining on a relatively long debate text.

[0054] In one example, the method is applied to the debate field. The first text may be a positive debate text, and the second text may be a negative debate text. Through the method, positive and negative views can be mined from the positive debate text and the negative debate text.

[0055] In another example, the method is applied to the intelligent justice field. The first text may be the text provided by the plaintiff (complaint), and the second text may be the text provided by the defendant (answer). Through the method, the disputed focus between the plaintiff and the defendant can be mined.

[0056] In yet another example, the method is applied to the field of e-commerce review information processing. The first text may be the product review text provided by the buyer user, and the second text may be the review reply text provided by the seller user. Through the method, focus mining can be performed from the communication between the buyer and the seller, and valuable questions and answers can be mined.

[0057] Figure 2Shows the argumentative text used in this embodiment, including a pair of opening arguments (the first text) and rebuttals (the second text). Here, the data pairs of review opinions (Review) and replies (Rebuttal) are used to verify the matching view recognition model. The method can perform matching view mining on relatively long argumentative texts, and for a given argumentative text, automatically extract each corresponding matching view in the two texts. As Figure 2 shown, if two sentences in the opening argument and the rebuttal are a matching view, the numbers after their labels can be the same, where REVIEW-1 (review opinion) and REPLY-1 (reply) represent the first matching view in these two paragraphs. Each argument here can be one sentence or multiple sentences, and there may be no overlapping parts between different arguments.

[0058] Step S103: Through the shared sentence vector determination module in the matching view recognition model, determine the shared sentence vector of the sentence according to the word vectors in the sentence.

[0059] The method divides the task of matching view mining into two subtasks: argument mining and argument pairing. The matching view recognition model adopts a multi-task training model framework, including: a shared sentence vector determination module, an argument classifier, and a matching view classifier. In this embodiment, the argument mining subtask is regarded as a sentence-level sequence labeling problem. Through the argument classifier, label whether each sentence is an argument, and the IOBES label can be used to distinguish the extracted argument and non-argument parts. In this embodiment, the argument pairing subtask is regarded as a sentence-level binary classification problem. Through the matching view classifier, each sentence in the first text and each sentence in the second text are paired pairwise and binary classified to determine whether the two sentences are an argument pair.

[0060] The first text and the second text respectively include multiple sentences. Before performing the corresponding classification processing through the argument classifier and the matching view classifier, it is necessary to determine the shared sentence vector of each sentence in the two texts through the shared sentence vector determination module according to the word vectors in the sentence. Since the matching view recognition model proposed in this embodiment does not separately learn two different sentence vector representations for the two subtasks (argument mining and matching view mining), but through a multi-task model, performs argument mining and matching view mining based on the same sentence vector, the sentence vector learned in this embodiment is a vector shared by the two subtasks, called the shared sentence vector. The shared sentence vector is related to the vector representation of the sentence itself and the vector representations of other sentences in the argumentative text. Among them, the vector of the sentence itself is related to the word vectors appearing in the sentence and has nothing to do with other sentences.

[0061] In specific implementation, step S103 can be implemented in the following manner: for each sentence in the text, determine the shared sentence vector of the sentence according to the word vector in the sentence and the argumentative text type of the sentence by matching the shared sentence vector determination module in the viewpoint recognition model.

[0062] In this embodiment, step S103 may include the following steps:

[0063] Step S1031: Determine the sentence vector according to the word vector in the sentence and the argumentative text type of the sentence through the sentence encoder in the shared sentence vector determination module.

[0064] The argumentative text type of a sentence can be an affirmative argument or a rhetorical question, and the argumentative text type of the sentence depends on whether the text to which the sentence belongs is an affirmative argument or a rhetorical question.

[0065] In this embodiment, the sentence vector of each sentence in the argumentative text is determined through the sentence encoder, that is, the vector of the sentence itself. The vector of the sentence itself is related to the word vectors that appear in the sentence and may also be related to the argumentative text type of the sentence, but is not related to other sentences. The sentence encoder can be a token-level long short-term memory network (Token-LSTM, abbreviated as T-LSTM), or other types of neural networks can be used, such as RNN, convolutional neural network, etc.

[0066] Figure 3 Shows a viewpoint recognition model based on a hierarchical long short-term memory structure (Hierarchical LSTM) proposed in this embodiment. Hierarchical long short-term memory is a structure that stacks word-level long short-term memory and sentence-level long short-term memory. Among them, the dashed box shows the sentence encoder. This encoder can use the word representations pre-trained by BERT as the input of the word-level long short-term memory to obtain the sentence representation (i.e., the sentence vector). Among them, t0…tT-1 represent tokens (each word in the sentence), x0…xT-1 are the representations of each word (i.e., word vectors), this sentence has a total of T words, hs is the vector representation of this sentence, and cs represents the argumentative text type, which is used to distinguish whether this sentence is an affirmative argument or a rhetorical question. Figure 3 It can be seen that the first text (affirmative argument) and the second text (rhetorical question) have a total of n sentences, and are the first and second sentences in the first text, represents the second-to-last sentence in the second text, represents the last sentence in the second text. Among them, the training of word vectors belongs to the relatively mature existing technology, so it will not be elaborated here.

[0067] Step S1033: Determine the shared sentence vector according to the sentence vectors of multiple sentences through the sentence vector sharing sub-module in the shared sentence vector determination module.

[0068] In this embodiment, for the sentences in the debate text, whether they appear in the first text or the second text, the sentence vector sharing sub-module determines the shared sentence vector of each sentence in the debate text according to the vector of the sentence itself and the vectors of other sentences (including the argument sentences and cross-examination sentences) in the debate text. Therefore, the shared sentence vector is related not only to the vector of the sentence itself but also to the vector representations of other sentences in the debate text.

[0069] As Figure 3 shown, the sentence vector sharing sub-module can adopt a long short-term memory model based on sentences, that is, a sentence-level long short-term memory model (Sentence-LSTM, abbreviated as S-LSTM). The sentence vector sharing sub-module can adopt an S-LSTM network or other types of neural networks, such as RNN, convolutional neural networks, etc. In specific implementation, the sentence representation output by the sentence encoder is input into the sentence-level long short-term memory (S-LSTM) to perform sequence annotation on each sentence of the entire debate text and generate the shared sentence vector of the sentence. Figure 3 The output of S-LSTM in

[0070] Step S205: By matching the argument classifier in the view recognition model, determine whether the sentence is an argument according to the shared sentence vector; and, combine the sentences in the first text and the sentences in the second text pairwise, and by matching the matching view classifier in the view recognition model, determine whether the two sentences are matching views according to the shared sentence vector.

[0071] In this embodiment, according to the shared sentence representation (i.e., the shared sentence vector) generated by the shared sentence vector determination module, two types of labels can be predicted simultaneously, that is, whether the sentence is an argument and whether two sentences are matching views.

[0072] As Figure 3 shown, in this embodiment, the argument classifier (which can adopt CRF, conditional random field) is used to label the results of argument mining, which is a sentence-level annotation result, and its output can be I / O / B / E / S. In the IOBES output by the argument classifier, O represents that the sentence is not an argument; if an argument consists of one sentence, it is labeled as S; if an argument consists of multiple sentences, it is labeled as B I…I E, B represents the first sentence of the argument, I is the middle sentence, and E is the last sentence. The matching view classifier (which can adopt a linear Linear model) is used to label whether two arguments are paired, and its output can be 1 or 0. If both of these two sentences are arguments and belong to the same matching view, they can be labeled as 1; otherwise, they are labeled as 0.

[0073] As can be seen from the above steps, the method provided by the embodiments of the present application performs matching view mining processing through a matching view recognition model obtained by a multi-task training method. The model has two training objectives, namely argument sequence annotation and matching view classification, and these two sub-tasks are trained together. The training data set uses a whole review opinion and its reply as a type of argumentative text to verify the model effect.

[0074] In one example, step S205 may include the following sub-steps:

[0075] Step S2051: For each sentence, use the argument classifier in the matching view recognition model to determine whether the sentence is an argument according to the shared sentence vector;

[0076] Step S2053: According to the type of argumentative text and the argument judgment result, determine the argument sentences of the first text and the second text that are paired in pairs; wherein, the type of argumentative text includes establishing an argument or asking a question.

[0077] Step S2055: For each pair combination of the argument sentences of the first text and the argument sentences of the second text, use the matching view classifier in the matching view recognition model to determine whether the two argument sentences are matching views.

[0078] Through steps S2051 to S2055, first use the sequence annotation result (shared sentence vector) output by the shared sentence vector determination module, mine out the arguments through the argument classifier, and then pair these arguments in pairs through the matching view classifier to see if they can form matching views. In this way, only the argument sentences need to be judged for matching views, which can effectively reduce the amount of calculation and thus improve the efficiency of matching view mining.

[0079] In specific implementation, it can also be to first use the matching view classifier in the matching view recognition model to determine whether two sentences match, and then for the matching sentences, use the argument classifier to determine whether they are argument sentences. If both sentences are arguments, it is determined that the two are matching views.

[0080] In one example, the method may further include the following steps: learning the matching view recognition model from the argumentative text and the matching view annotation information set. In specific implementation, each piece of training data may include two types of information: the first text and the second text for training, and the matching view annotation information between the two texts (such as which sentences are matching views), and the tags as shown in Figure 2 may be used.

[0081] As can be seen from the above description, the matching view recognition model adopted by the method is a multi-task trained matching view recognition model. During the prediction process, the trained multi-task model can be decomposed into two sub-modules to perform two sub-tasks in a pipeline manner to extract the final matching view. For example, first, the argumentation type of a sentence is determined through the argumentation recognition sub-module (including the shared sentence vector determination module and the argumentation classifier), and then the matching view is determined through the matching view recognition sub-module (including the argumentation classifier). This multi-task training model utilizes the complementary content and structural relationship of question and answer in argumentative texts. Compared with the pipeline structure of separately training the argumentation recognition model and the matching view recognition model, it can learn better shared sentence representations, thereby achieving better matching view mining results.

[0082] As can be seen from the above embodiments, the matching view mining method provided by the embodiments of the present application includes obtaining a first text and a second text including matching views; for each sentence in the text, through the shared sentence vector determination module in the matching view recognition model, determining the shared sentence vector of the sentence according to the word vectors in the sentence; through the argumentation classifier in the matching view recognition model, judging whether the sentence is an argument according to the shared sentence vector; and for each pair of sentences, through the matching view classifier in the matching view recognition model, judging whether the two sentences are matching views according to the shared sentence vector. This processing method splits the matching view mining task into two sub-tasks of argumentation mining and argumentation pairing, and effectively combines these two sub-tasks through multi-task training of the matching view recognition model, utilizes the complementary content and structural relationship of question and answer in argumentative texts, learns better shared sentence representations, and then mines the corresponding matching views in the argumentative text; therefore, the accuracy of matching view mining can be effectively improved.

[0083] Second Embodiment

[0084] In the above embodiments, a matching view mining method is provided. Correspondingly, the present application also provides a matching view mining device. This device corresponds to the embodiments of the above method. The parts of this embodiment that are the same as those in the first embodiment will not be described again. Please refer to the corresponding parts in the first embodiment.

[0085] A matching view mining device provided by the present application includes:

[0086] An argumentative text acquisition unit, configured to acquire a first text and a second text including matching views, where the first text and the second text include sentences;

[0087] A shared sentence vector determination unit, configured to determine the shared sentence vector of a sentence according to the word vectors in the sentence through the shared sentence vector determination module in the matching view recognition model;

[0088] A matching view determination unit, configured to determine whether a sentence is an argument according to the shared sentence vector by means of an argument classification classifier in a matching view recognition model; and, pair up sentences in the first text with sentences in the second text, and determine whether the two sentences are matching views according to the shared sentence vector by means of a matching view classification classifier in the matching view recognition model.

[0089] In one example, the shared sentence vector determination unit may include:

[0090] A sentence encoder, configured to determine a sentence vector according to word vectors in a sentence and the argument text type of the sentence;

[0091] A sentence vector sharing sub-module, configured to determine a shared sentence vector according to sentence vectors of multiple sentences.

[0092] In one example, the matching view determination unit may include:

[0093] An argument classification classifier, configured to determine whether a sentence is an argument according to the shared sentence vector;

[0094] A pairing sub-unit, configured to determine argument sentences in the first text and argument sentences in the second text that are paired up according to the argument text type of the sentence and the argument judgment result;

[0095] A matching view classification classifier, configured to determine whether two argument sentences are matching views according to the shared sentence vector.

[0096] In one example, the apparatus may further include:

[0097] A model construction unit, configured to learn a matching view recognition model from an argument text and a matching view annotation information set.

[0098] In one example, the apparatus may further include:

[0099] The word vectors are obtained through BERT pre-training.

[0100] In one example, the first text includes: a proponent's argument text, and the second text includes: an opponent's argument text.

[0101] In one example, the first text includes: text provided by a plaintiff, and the second text includes: text provided by a defendant.

[0102] In one example, the first text includes: product review text provided by a buyer user, and the second text includes: review reply text provided by a seller user.

[0103] The third embodiment

[0104] The present application also provides an electronic device. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple. For related parts, please refer to the partial description of the method embodiments. The device embodiments described below are only illustrative.

[0105] An electronic device according to this embodiment, the electronic device includes: a processor and a memory; the memory is used to store a program for implementing the matching view mining method. After the device is powered on and runs the program of this method through the processor, the following steps are executed: obtaining a first text and a second text including matching views, where the first text and the second text include sentences; through the shared sentence vector determination module in the matching view recognition model, determining the shared sentence vector of the sentence according to the word vectors in the sentence; through the argument classification classifier in the matching view recognition model, judging whether the sentence is an argument according to the shared sentence vector; and, combining the sentences in the first text and the sentences in the second text pairwise, and through the matching view classification classifier in the matching view recognition model, determining whether the two sentences are matching views according to the shared sentence vector.

[0106] The fourth embodiment

[0107] Please refer to Figure 4 , which is a schematic flowchart of the embodiment of the method for constructing the matching view recognition model of the present application. The execution subject of the method includes but is not limited to a server, and can also be any device capable of implementing the method. In this embodiment, the method may include the following steps:

[0108] Step S401: Construct the network structure of the matching view recognition model, and the classification model includes a shared sentence vector determination module, an argument classification classifier, and a matching view classification classifier.

[0109] In this embodiment, the shared sentence vector determination module is used to determine the shared sentence vector of the sentence according to the word vectors in the sentence. The argument classification classifier is used to judge whether the sentence is an argument according to the shared sentence vector. The matching view classification classifier is used to judge whether two sentences are matching views according to the shared sentence vector.

[0110] Step S403: Learn the parameters of the matching view recognition model from the argumentative text and the matching view annotation information set.

[0111] During the process of training the model, two loss functions may be included, one is the loss function for argument classification, and the other is the loss function for matching view classification.

[0112] As can be seen from the above embodiments, the method for constructing a matching view recognition model provided by the embodiments of the present application constructs the network structure of the matching view recognition model. The model includes a shared sentence vector determination module, an argument classifier, and a matching view classifier, and learns the parameters of the matching view recognition model from the argumentative text and the matching view annotation information set. This processing method enables the matching view recognition model to be trained through multi-tasking, effectively combines the two subtasks of argument mining and argument pairing, and utilizes the complementary content and structural relationship of question-and-answer in the argumentative text to learn a better shared sentence representation, and then mines the corresponding matching views in the argumentative text. Therefore, the accuracy of the model can be effectively improved.

[0113] The fifth embodiment

[0114] In the above embodiment, a method for constructing a matching view recognition model is provided. Correspondingly, the present application also provides a device for constructing a matching view recognition model. This device corresponds to the embodiment of the above method. The parts of this embodiment that are the same as those in the fourth embodiment will not be described in detail. Please refer to the corresponding parts in the fourth embodiment.

[0115] A device for constructing a matching view recognition model provided by the present application includes:

[0116] A network construction unit, configured to construct the network structure of the matching view recognition model. The model includes a shared sentence vector determination module, an argument classifier, and a matching view classifier. The shared sentence vector determination module is configured to determine the shared sentence vector of a sentence according to the word vectors in the sentence. The argument classifier is configured to determine whether a sentence is an argument according to the shared sentence vector. The matching view classifier is configured to determine whether two sentences are matching views according to the shared sentence vector.

[0117] A training unit, configured to learn the parameters of the matching view recognition model from the argumentative text and the matching view annotation information set.

[0118] The sixth embodiment

[0119] The present application also provides an electronic device. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For related parts, please refer to the partial description of the method embodiment. The device embodiment described below is only illustrative.

[0120] An electronic device according to this embodiment, the electronic device includes: a processor and a memory; the memory is used to store a program for implementing the method of constructing a matching view recognition model. After the device is powered on and the program of this method runs through the processor, the following steps are executed: constructing a network structure of the matching view recognition model, the model includes a shared sentence vector determination module, an argument classifier, and a matching view classifier; the shared sentence vector determination module is used to determine the shared sentence vector of the sentence according to the word vectors in the sentence; the argument classifier is used to judge whether the sentence is an argument according to the shared sentence vector; the matching view classifier is used to judge whether two sentences are matching views according to the shared sentence vector; learning the parameters of the matching view recognition model from the argumentative text and the matching view annotation information set.

[0121] The seventh embodiment

[0122] This application also provides a method for mining matching views. The execution subject of the method includes but is not limited to a server, and can also be any device capable of implementing the method. In this embodiment, the method may include the following steps:

[0123] Step S501: Obtain a first text and a second text including matching views, and the first text and the second text include sentences.

[0124] Step S503: Through an argument recognition model, judge whether a sentence is an argument according to the word vectors in the sentence.

[0125] The argument recognition model can adopt an existing argument recognition model. Since the argument recognition model belongs to relatively mature existing technology, it will not be elaborated here.

[0126] Step S505: Combine the argument sentences in the first text and the argument sentences in the second text pairwise, and through a matching view recognition model, judge whether two argument sentences are matching views according to the word vectors in the sentences.

[0127] The matching view recognition model may include a sentence encoder and a classifier. The sentence encoder can be used to determine the sentence vector of each sentence in the argumentative text. This sentence vector is related to the recognition of matching views and has nothing to do with argument mining. It is not a shared sentence vector.

[0128] The difference between the method provided in this embodiment and the method provided in Embodiment 1 is that: the method provided in this embodiment trains the argument recognition model and the matching view recognition model separately, combines the argument mining subtask and the argument pair mining subtask through a pipeline structure, and does not learn the shared sentence representation of the two tasks. Therefore, the effect of mining matching views is worse than the method provided in Embodiment 1.

[0129] As can be seen from the above embodiments, the matching view mining method provided by the embodiments of the present application obtains a first text and a second text including matching views; for each sentence in the text, through an argumentation recognition model, according to the word vectors in the sentence, it is determined whether the sentence is an argument; for each pair combination of the argument sentences in the first text and the argument sentences in the second text, through a matching view recognition model, according to the word vectors in the sentence, it is determined whether the two argument sentences are matching views; this processing method splits the matching view mining task into two subtasks of argumentation mining and argumentation pairing, trains two models independently, combines these two models through a pipeline structure, and utilizes the complementary content and structural relationship of the question-and-answer in the argumentative text to mine the corresponding matching views in the argumentative text; therefore, the effect of matching view mining can be effectively improved.

[0130] The eighth embodiment

[0131] In the above embodiment, a matching view mining method is provided. Correspondingly, the present application also provides a matching view mining device. This device corresponds to the embodiment of the above method. The parts that are the same as those in the seventh embodiment will not be described in detail. Please refer to the corresponding parts in the seventh embodiment.

[0132] A matching view mining device provided by the present application includes:

[0133] An argumentative text acquisition unit, configured to acquire a first text and a second text including matching views, where the first text and the second text include sentences;

[0134] An argumentation recognition unit, configured to determine whether a sentence is an argument according to the word vectors in the sentence through an argumentation recognition model;

[0135] A matching view recognition unit, configured to combine the argument sentences in the first text and the argument sentences in the second text in pairs, and determine whether the two argument sentences are matching views according to the word vectors in the sentence through a matching view recognition model.

[0136] The ninth embodiment

[0137] The present application also provides an electronic device. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple. For the relevant parts, please refer to the partial description of the method embodiment. The device embodiment described below is only illustrative.

[0138] An electronic device according to this embodiment, the electronic device includes: a processor and a memory; the memory is used to store a program for implementing the matching view mining method. After the device is powered on and runs the program of this method through the processor, the following steps are executed: obtaining a first text and a second text including matching views, the first text and the second text including sentences; through an argumentation recognition model, judging whether a sentence is an argument according to the word vectors in the sentence; combining the argument sentences in the first text with the argument sentences in the second text pairwise, and through a matching view recognition model, judging whether two argument sentences are matching views according to the word vectors in the sentence.

[0139] Tenth Embodiment

[0140] In the above embodiment, a matching view mining method is provided. Correspondingly, the present application also provides a debate robot system. This system corresponds to the embodiment of the above method. The parts of this embodiment that are the same as those in the first embodiment will not be described in detail again. Please refer to the corresponding parts in the first embodiment.

[0141] A debate robot system provided by the present application includes: a matching view mining device and a debate device based on matching views. The system can extract corresponding pro and con arguments from a large number of documents through the matching view mining device; and use the extracted matching views for subsequent pro and con speeches through the debate device based on matching views.

[0142] Among them, the matching view mining device is used to obtain a first text and a second text including matching views, the first text and the second text including sentences; determine the shared sentence vector of the sentence according to the word vector in the sentence through the shared sentence vector determination module in the matching view recognition model; judge whether the sentence is an argument according to the shared sentence vector through the argumentation classifier in the matching view recognition model; and combine the sentences in the first text with the sentences in the second text pairwise, and determine whether two sentences are matching views according to the shared sentence vector through the matching view classifier in the matching view recognition model.

[0143] As can be seen from the above embodiments, the debate robot system provided by the embodiments of the present application includes a matching view mining device and a debate device based on matching views. For a large number of historical debate texts, the matching view mining device determines the shared sentence vector of a sentence according to the word vectors in the debate text sentence through the shared sentence vector determination module in the matching view recognition model; determines whether the sentence is an argument according to the shared sentence vector through the argument classification in the matching view recognition model; and combines the argument sentences and the interrogation sentences in pairs, and determines whether the two sentences are matching views according to the shared sentence vector through the matching view classifier in the matching view recognition model; then, through the debate device based on matching views, the matching view mining result is used for the subsequent speeches of the affirmative and negative sides; this processing method splits the matching view mining task into two subtasks of argument mining and argument pairing, and effectively combines these two subtasks by training the matching view recognition model with multiple tasks, and uses the complementary content and structural relationship of the question-and-answer in the debate text to learn a better shared sentence representation, and then mines the one-to-one matching views in the debate text; therefore, the accuracy of matching view mining can be effectively improved, and then the accuracy of the automatic speech of the debate robot can be improved.

[0144] Although the present application is disclosed above with preferred embodiments, it is not intended to limit the present application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the protection scope of the present application should be subject to the scope defined by the claims of the present application.

[0145] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0146] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of, for example, read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0147] 1. A computer-readable medium includes permanent and non-permanent, removable and non-removable media that can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, 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, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information accessible by a computing device. As defined herein, a computer-readable medium does not include transitory media such as modulated data signals and carrier waves.

[0148] 2. Those skilled in the art will appreciate that the embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

Claims

1. A method for mining matching viewpoints, characterized in that, include: Obtaining a first text and a second text including matching viewpoints, wherein the first text and the second text include sentences; By matching the shared sentence vector determination module in the opinion recognition model, a shared sentence vector of the sentence is determined based on the word vectors in the sentence, where the shared sentence vector is related to the vector representation of the sentence itself and the vector representations of other sentences in the debate text; By using the debate classifier in the matching viewpoint recognition model, based on the shared sentence vector, it is determined whether the sentence is an argument; and, sentences in the first text are combined with sentences in the second text in pairs, and by using the matching viewpoint classifier in the matching viewpoint recognition model, based on the shared sentence vector, it is determined whether the two sentences are matching viewpoints.

2. The method according to claim 1, wherein The shared sentence vector determination module in the matching opinion recognition model determines the shared sentence vector of the sentence based on the word vectors in the sentence, including: The sentence encoder in the shared sentence vector determination module determines the sentence vector based on the word vectors in the sentence and the argument text type of the sentence; The sentence vector sharing submodule in the shared sentence vector determination module determines a shared sentence vector based on the sentence vectors of multiple sentences.

3. The method according to claim 1, characterized in that By matching the argument classifier in the opinion recognition model, we can determine whether a sentence is an argument based on the shared sentence vector. According to the argument text type and argument judgment results of the sentences, the argument sentences of the first text and the argument sentences of the second text are determined in pairs; The matching opinion classifier in the matching opinion identification model is used to determine whether two argument sentences have matching opinions.

4. The method according to claim 1, wherein Also includes: The matching opinion recognition model is learned from the debate text and the matching opinion annotation information set.

5. The method according to claim 1, wherein Also includes: The word vector is obtained through BERT pre-training.

6. The method according to claim 1, characterized in that The first text includes: the affirmative side debate text, and the second text includes: the negative side debate text.

7. The method according to claim 1, characterized in that The first text includes: a text provided by the plaintiff, and the second text includes: a text provided by the defendant.

8. The method according to claim 1, characterized in that The first text includes: a product review text provided by a buyer user, and the second text includes: a review reply text provided by a seller user.

9. A method for constructing a matching view recognition model, characterized in that include: Constructing a network structure for a matching viewpoint identification model, the model includes a shared sentence vector determination module, an argument classifier, and a matching viewpoint classifier; the shared sentence vector determination module is used to determine a sentence's shared sentence vector based on word vectors in the sentence, wherein the shared sentence vector is related to the vector representation of the sentence itself and the vector representations of other sentences in the debate text; the argument classifier is used to determine whether a sentence is an argument based on the shared sentence vector; and the matching viewpoint classifier is used to determine whether two sentences have matching viewpoints based on the shared sentence vector. The parameters of the matching opinion recognition model are learned from the debate text and the matching opinion annotation information set.

10. A matching opinion mining device, characterized in that, include: An argumentative text acquisition unit for acquiring a first text and a second text including matching viewpoints, where the first text and the second text include sentences; A shared sentence vector determination unit for determining a shared sentence vector of a sentence through a shared sentence vector determination module in an argument identification model, based on the word vectors in the sentence, where the shared sentence vector is related to the vector representation of the sentence itself and the vector representations of other sentences in the argumentative text; An argument matching determination unit for determining whether a sentence is an argument through an argument classification classifier in an argument identification model, based on the shared sentence vector; and for pairwise combining the sentences in the first text with the sentences in the second text, and determining whether two sentences are matching viewpoints through a matching viewpoint classification classifier in the argument identification model, based on the shared sentence vector.

11. A debate robot system, characterized in that, Comprising: The matching viewpoint mining device according to claim 10 above, and an argumentation device based on matching viewpoints.

Images