Intelligent comment reply method and device

Abstract

The present application provides a method and device for intelligent comment reply, wherein the intelligent comment reply method includes: receiving user comments on a target object, and inputting the text information contained in the user comments and the corpus corresponding to the target object into a first feature extraction model to obtain the emotion category and comment category in the text information, wherein the corpus contains at least one corpus information; screening at least one initial reply content matching the emotion category and the comment category from a historical database; inputting the at least one initial reply content into a second feature extraction model to obtain a target reply content; and replying to the comment to be replied in an interactive page based on the target reply content. Through the method provided by the present application, a neural network model can be used to automatically generate a reply text for a target object, thereby improving the efficiency of the reply.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and more particularly to an intelligent comment reply method, an intelligent comment reply device, a computing device, and a computer-readable storage medium. Background Art

[0002] With the development of network technology, there are more and more network technology products. Correspondingly, when these products are facing users, users who use the products are also required to give feedback on the products. Therefore, there are more and more user comments on various products.

[0003] In the prior art, user comments are generally processed by manual replies, and the staff who reply are called customer service staff. However, the number and energy of customer service staff are limited. Manually replying to user comments will not only take a lot of time, but also cause the number of users waiting for replies to gradually increase, further causing users to wait, creating a vicious cycle; customer service staff will also feel bored because they reply to a large number of repetitive user comments, resulting in low work efficiency. Summary of the invention

[0004] In view of this, an embodiment of the present application provides an intelligent comment reply method. The present application also relates to an intelligent comment reply device, a computing device, and a computer-readable storage medium to solve the defect of low efficiency of manual comment reply in the prior art.

[0005] According to a first aspect of an embodiment of the present application, there is provided a smart comment reply method, comprising:

[0006] Receiving user comments on a target object, and inputting text information contained in the user comments and a corpus corresponding to the target object into a first feature extraction model to obtain an emotion category and a comment category in the text information, wherein the corpus contains at least one piece of corpus information;

[0007] Filtering at least one initial reply content matching the emotion category and the comment category from a history database;

[0008] Inputting the at least one initial reply content into a second feature extraction model to obtain target reply content;

[0009] Reply to the comment to be replied in the interactive page based on the target reply content.

[0010] According to a second aspect of an embodiment of the present application, a smart comment reply method is provided, comprising:

[0011] Displaying a comment input page for the user according to the user's call request, and receiving a user comment to be replied to input by the user based on the comment input page;

[0012] Inputting the text information in the user comments to be replied to into a first feature extraction model to obtain the emotion category and comment category in the text information;

[0013] Filtering at least one initial reply content matching the emotion category and the comment category from a history database;

[0014] Inputting the at least one initial reply content into a second feature extraction model to obtain target reply content;

[0015] Reply to the comment to be replied in the interactive page based on the target reply content.

[0016] According to a third aspect of an embodiment of the present application, there is provided an intelligent comment replying device, comprising:

[0017] A first receiving module is configured to receive user comments on a target object, and input text information contained in the user comments and a corpus corresponding to the target object into a first feature extraction model to obtain an emotion category and a comment category in the text information, wherein the corpus contains at least one piece of corpus information;

[0018] A first screening module is configured to screen at least one initial reply content matching the emotion category and the comment category from a history database;

[0019] A first acquisition module is configured to input the at least one initial reply content into a second feature extraction model to obtain a target reply content;

[0020] The first reply module is configured to reply to the comment to be replied in the interactive page based on the target reply content.

[0021] According to a fourth aspect of an embodiment of the present application, there is provided an intelligent comment replying device, comprising:

[0022] A second receiving module is configured to display a comment input page for the user according to a call request of the user, and receive a user comment to be replied input by the user based on the comment input page;

[0023] An input module, configured to input the text information in the user comment to be replied to into a first feature extraction model to obtain the emotion category and comment category in the text information;

[0024] A second screening module is configured to screen at least one initial reply content matching the emotion category and the comment category from the history database;

[0025] A second input module is configured to input the at least one initial reply content into a second feature extraction model to obtain a target reply content;

[0026] The second reply module is configured to reply to the comment to be replied in the interactive page based on the target reply content.

[0027] According to a fifth aspect of an embodiment of the present application, a first feature extraction model training device is provided, comprising:

[0028] A sample acquisition module is configured to acquire sample comments, category labels corresponding to the sample comments, and a corpus, wherein the sample comments are comment information of a target object, the corpus is at least one piece of corpus information corresponding to the target object, and the category labels include emotion category labels and comment category labels;

[0029] A processing module, configured to input the sample comment and the at least one piece of corpus information into a first feature extraction model, and obtain a predicted emotion category and a predicted comment category output by the first feature extraction model;

[0030] a calculation module, configured to calculate a loss value of the first feature extraction model according to the emotion category label, the predicted emotion category, the comment category label and the predicted comment category;

[0031] A training module is configured to train the first feature extraction model based on the loss value until a training stop condition is reached.

[0032] According to the sixth aspect of an embodiment of the present application, a computing device is provided, comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, wherein the processor implements the steps of the intelligent comment reply method when executing the instructions.

[0033] According to the seventh aspect of the embodiments of the present application, a computer-readable storage medium is provided, which stores computer instructions, and when the instructions are executed by a processor, the steps of the intelligent comment reply method are implemented.

[0034] The present application provides an intelligent comment reply method, which receives user comments on a target object, and inputs text information contained in the user comments and a corpus corresponding to the target object into a first feature extraction model to obtain an emotion category and a comment category in the text information, wherein the corpus contains at least one corpus information; at least one initial reply content matching the emotion category and the comment category is filtered out from a historical database; the at least one initial reply content is input into a second feature extraction model to obtain a target reply content; and based on the target reply content, the comment to be replied to is replied to in an interactive page.

[0035] An embodiment of the present application realizes automatic generation of reply text for the target object through a neural network model, thereby improving the efficiency of reply. Moreover, the existing user comments can be effectively covered through the corpus and historical database, providing a basis for the model to automatically generate reply text. In addition, replying based on the automatically generated text in the interactive page also ensures the efficiency of processing user comments and improves work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flowchart of an intelligent comment reply method provided by an embodiment of the present application;

[0037] Figure 2 is a flow chart of a first feature extraction model provided in an embodiment of the present application;

[0038] Figure 3 This is a flowchart of an intelligent comment reply method provided by an embodiment of the present application;

[0039] Figure 4 It is a schematic diagram of the review process provided by an embodiment of the present application;

[0040] Figure 5 This is a schematic diagram of a specific application process of an intelligent comment reply method provided by an embodiment of the present application;

[0041] Figure 6 This is a flowchart of an intelligent comment reply method provided by an embodiment of the present application;

[0042] Figure 7 This is a schematic diagram of the application effect of an intelligent comment reply method provided by an embodiment of the present application;

[0043] Figure 8 It is a schematic diagram of a Bert model structure provided in an embodiment of the present application;

[0044] Fig. 9 This is a schematic diagram of the application process of an intelligent comment reply method provided by an embodiment of the present application;

[0045] Fig.10 is a schematic diagram of the training process of the first feature extraction model provided in one embodiment of the present application;

[0046] Fig.11 It is a structural diagram of an intelligent comment replying device provided by an embodiment of the present application;

[0047] Fig.12 It is a structural diagram of an intelligent comment replying device provided by an embodiment of the present application;

[0048] Fig.13 It is a structural schematic diagram of a training device for a first feature extraction model provided in one embodiment of the present application;

[0049] Fig.14 It is a structural block diagram of a computing device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0050] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0051] The terms used in one or more embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of the present application. The singular forms of "a", "said" and "the" used in one or more embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of the present application refers to and includes any or all possible combinations of one or more associated listed items.

[0052] It should be understood that, although the terms first, second, etc. may be used to describe various information in one or more embodiments of the present application, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0053] First, the terms involved in one or more embodiments of the present application are explained.

[0054] The first feature extraction model is a model obtained by improving the Bert model in the prior art. Compared with the existing Bert model, a voting layer is added.

[0055] The second feature extraction model: pointer generation network model.

[0056] Iteration: The activity of repeating a feedback process, usually with the goal of moving closer to a desired goal or outcome.

[0057] Encapsulation: refers to packaging different parts of a program into a complete, callable program.

[0058] In the present application, an intelligent comment reply method is provided. The present application also relates to an intelligent comment reply device, a computing device, and a computer-readable storage medium, which are described in detail one by one in the following embodiments.

[0059] Figure 1 A flowchart of the steps of an intelligent comment method provided according to an embodiment of the present application is shown, specifically including steps 102 to 108.

[0060] Step 102: Receive user comments on a target object, and input text information contained in the user comments and a corpus corresponding to the target object into a first feature extraction model to obtain emotion categories and comment categories in the text information, wherein the corpus contains at least one piece of corpus information.

[0061] In actual applications, responses to user comments are generally handled by human labor, that is, manual customer service. In a few cases, the method of applying templates and adding keywords to respond to user comments is also used. However, the response method of manual customer service depends on the customer service staff who respond. When faced with a large number of user comments, customer service staff will face the problems of insufficient manpower, inability to handle user comments in a timely manner, and being annoyed by a large number of repeated comments. Although the method of applying templates and adding keywords is highly versatile, it cannot make targeted responses to a specific product, and the method of applying templates and keywords will produce a large number of repeated responses, which will also cause user dissatisfaction.

[0062] Therefore, in order to save the cost of manual replies and improve reply efficiency, the present application provides an intelligent comment reply method, which can receive user comments on a target object, and input the text information contained in the user comments and the corpus corresponding to the target object into a first feature extraction model to obtain the emotion category and comment category in the text information, and filter at least one initial reply text that matches the emotion category and the comment category from the historical data block, and input the at least one initial reply text into a second feature extraction model to obtain the target content, and reply to the comment to be replied in the interactive page based on the target content. Through the above method, a neural network model is used to automatically generate reply texts for the target object, which saves labor costs, effectively improves the reply speed and efficiency of manual customer service, and also improves the specificity for different target objects.

[0063] In step 102, the first feature extraction model can be trained by the following method: Figure 2 As shown, the method comprises the following steps:

[0064] Step 202: Obtain sample comments, category labels corresponding to the sample comments, and a corpus, wherein the sample comments are comment information of a target object, the corpus is at least one piece of corpus information corresponding to the target object, and the category labels include emotion category labels and comment category labels.

[0065] Step 204: Input the sample comment and the at least one piece of corpus information into a first feature extraction model to obtain a predicted emotion category and a predicted comment category output by the first feature extraction model.

[0066] Step 206: Calculate the loss value of the first feature extraction model according to the emotion category label, the predicted emotion category, the comment category label and the predicted comment category.

[0067] Step 208: Train the first feature extraction model based on the loss value until a training stop condition is reached.

[0068] The training method of the first feature extraction model provided in the embodiment of the present application obtains sample comments, category labels and corpora corresponding to the sample comments, and inputs the sample comments, category labels and corpora corresponding to the sample comments into the first feature extraction model, so that the first feature extraction model can understand user comments more comprehensively and accurately; then the first feature extraction model is iteratively trained according to the predicted emotion category labels, the predicted category labels and emotion labels, and the category labels. On the basis of the comprehensive and accurate understanding of user comments, the first feature extraction model can further analyze the user comments, thereby improving the accuracy of the replies and improving the effect of model training.

[0069] In an optional implementation of this embodiment, the following further includes:

[0070] The target object includes at least one of a game, an audio, a commodity, an article, and a video.

[0071] It should be noted that the target object can be a game, audio, video, etc., and this solution does not impose any restrictions on this. For example, if the target object is a game, the user comments on the target object can be player comments on the game. Alternatively, if the target object is audio, such as music, the user comments on the target object can be listener comments on the song. The interactive page can be a comment area for the target object. Continuing with the above example, if the target object is music, the interactive page can be a comment area for the application used to play the music. The intelligent comment reply method in this solution can be applied to corresponding target scenarios for different target objects, thereby improving applicability.

[0072] like Figure 3 As shown, Figure 3 This is a flow chart of the first feature extraction model of this embodiment, wherein A, B, C, and D represent four different comment categories obtained after the first feature extraction model performs feature extraction on user comments. For example, if A corresponds to a certain game, it can be determined that the user comment is a comment on the game.

[0073] In an optional implementation of this embodiment, preprocessing the text information contained in the user comment includes:

[0074] The text information is preprocessed by removing duplicates and clearing garbled information.

[0075] In practical applications, before inputting the received user comments into the first feature extraction model, the user comments need to be preprocessed. The preprocessing methods for user comments include but are not limited to: deduplication, long paragraph sorting, spam removal, garbled code removal, and hyperlink removal. By preprocessing the received user comments, it is ensured that the text data input into the first feature extraction model are all usable data, and it is also ensured that the first feature extraction model can accurately identify.

[0076] In addition, the corpus corresponds to the target object, and may be a collection of professional terms related to the target object summarized from various forums, encyclopedias and other channels.

[0077] In an optional implementation of this embodiment, the text information contained in the user comment to be replied and the corpus corresponding to the target object are input into the first feature extraction model to obtain the emotion category and comment category in the text information, including:

[0078] Inputting the text information and the corpus into an embedding layer in a first feature extraction model, and obtaining a text information vector of the text information and at least one corpus information vector of the corpus through the embedding layer;

[0079] Inputting the text information vector and the at least one corpus information vector into an encoding layer in a first feature extraction model, and obtaining a text information encoding vector of the text information vector and at least one corpus information encoding vector of the at least one corpus information vector through the encoding layer;

[0080] The emotion category and the comment category of the text information are obtained according to the text information encoding vector and the at least one corpus information encoding vector.

[0081] For example, if the comment sent by the user is "@I feel the plot is not bad.www", the text information obtained through preprocessing is "I feel the plot is not bad", and "I feel the plot is not bad" and the corpus corresponding to this comment are input into the first feature extraction model. It should be noted that the first feature extraction model includes an embedding layer and an encoding layer.

[0082] By inputting user comments into the neural network model to obtain the sentiment labels and category labels of user comments, it is ensured that user comments can be classified specifically and clearly, thereby improving the efficiency of replying.

[0083] In an optional implementation of this embodiment, obtaining the text information vector of the text information and at least one corpus information vector of the corpus through the embedding layer includes:

[0084] Performing word segmentation processing on the text information vector and the at least one corpus information vector through the embedding layer to obtain word units of the text information and word units of the at least one corpus information;

[0085] Embedding the word units of the text information and the word units of the at least one piece of corpus information respectively to obtain the word vector, word vector and position vector of the text information and the word vector, word vector and position vector of the at least one piece of corpus information;

[0086] Summing the character vector, word vector and position vector of the text information to obtain a vector representation of the text information;

[0087] The character vector, word vector and position vector of the at least one piece of corpus information are summed to obtain a vector representation of the at least one piece of corpus information.

[0088] Continuing with the above example, the comment "I feel the plot is pretty good" and the corresponding corpus are input into the embedding layer in the first feature extraction model, and the comment "I feel the plot is pretty good" and the corresponding corpus are embedded. In the following, "I feel the plot is pretty good" is recorded as comment P1. For example, the corpus contains 3 pieces of corpus information, namely A1, A2 and A3. After embedding, the character vectors, word vectors and position vectors of P1, A1, A2 and A3 are obtained respectively. Then the character vectors, word vectors and position vectors of P1 are summed to obtain the vector representation of P1, and the character vectors, word vectors and position vectors of A1, A2 and A3 are summed to obtain the vector representation of A1, A2 and A3.

[0089] The vector representation of the text information and the corpus is obtained through the embedding layer, which ensures that the first feature extraction model can further process the user comments and also ensures that the text information to be processed will not be disordered.

[0090] In an optional implementation of this embodiment, obtaining the emotion category and the comment category of the text information according to the text information encoding vector and the at least one corpus information encoding vector includes:

[0091] Performing part-of-speech tagging on the text information encoding vector through the encoding layer to obtain the emotion category of the text information;

[0092] Through the encoding layer, part-of-speech tagging is performed on the text information encoding vector based on the at least one corpus information encoding vector to obtain a comment category of the text information.

[0093] Using the above example, P1 is tagged with part of speech through the encoding layer of the first feature extraction model. According to the information in "I feel the plot is pretty good", the emotion category of P1 can be obtained as: satisfied, the comment category is "game progress", and the emotion category also includes the corresponding emotion tendency, which can be represented by a value of 0-100. The closer the value is to 0, the more negative the emotion is, and the closer the value is to 100, the more positive the emotion is.

[0094] In addition, the corpus information A1, A2, and A3 in the corpus carry corresponding category labels. For example, if A1 contains the information "pass a level", the label of A1 is determined to be "game progress". The comment category of P1 is determined based on the vector representation of A1, A2, and A3 and the labels they carry.

[0095] Part-of-speech tagging is performed through the coding layer, which ensures that after the user comments are input into the first feature extraction model, the corresponding emotion category and comment category can be obtained, thereby ensuring the recognition accuracy of the first feature extraction model.

[0096] In an optional implementation of this embodiment, the first feature extraction model further includes a voting layer;

[0097] Inputting text information contained in the user comments to be replied to into a first feature extraction model to obtain the emotion category and comment category in the text information, including:

[0098] Inputting the text information into an embedding layer in a first feature extraction model, and obtaining a first vector representation of the text information through the embedding layer;

[0099] Processing the first vector representation of the text information through the encoding layer to obtain a first emotion category and a first comment category of the text information;

[0100] Inputting the first emotion category into a voting layer to obtain the emotion category voting result;

[0101] Inputting the emotion category voting result and the text information into an embedding layer in a first feature extraction model, and obtaining a second vector representation of the text information through the embedding layer;

[0102] The second vector representation of the text information is processed through the encoding layer to obtain a second emotion category and a second comment category of the text information and use the second emotion category and the second comment category as the emotion category and the comment category.

[0103] Continuing with the above example, after the first feature extraction model outputs P1 "I feel the plot is pretty good" and the corresponding emotion label and category label, the output P1 and the corresponding emotion label and category label are re-input into the first feature extraction model for iterative processing. Assuming that the current emotion label corresponding to P1 is "satisfied" and the corresponding value is 55, after being reprocessed by the voting layer, the corresponding value will be updated to 57, which is closer to the emotion in the user's comment. Moreover, the above re-input iteration process can be performed multiple times, and each output is closer to the user's true emotion than the previous one.

[0104] Reprocessing through the voting layer ensures that each output is closer to the user's real comment, further ensuring the accuracy of the first feature extraction model recognition.

[0105] Step 104: Filter at least one initial reply content matching the emotion category and the comment category from the history database.

[0106] In practical applications, the historical database may be composed of manually written reply texts, or reply information obtained from various channels such as forums and comment areas. Moreover, the reply texts in the historical database carry emotion tags and category tags. For example, a reply text in the historical database is "Thank you for your support, we will continue to work hard", which is hereinafter recorded as H1. Among them, the label carried by H1 is "for positive emotions, for game progress". According to the emotion category and the comment category, the label of the reply text in the historical database is matched to obtain at least one corresponding reply text. For example, the emotion category of the above-mentioned P1 is "satisfied", and the comment type is "game progress". According to "satisfied" and "game progress", at least one corresponding reply text "Thank you for your support, we will continue to work hard" is obtained in the historical database.

[0107] By matching the sentiment and comment categories with the corresponding tags in the historical database, it is ensured that the obtained reply text can accurately correspond to the user's pending comment in terms of sentiment and type, further increasing the accuracy of the reply.

[0108] Step 106: Input the at least one initial reply content into a second feature extraction model to obtain target reply content.

[0109] The second feature extraction model is a pointer generation network. Specifically, the second feature extraction model includes an encoder, a memory network and a decoder.

[0110] In an optional implementation of this embodiment, inputting the at least one initial reply content into a second feature extraction model to obtain target reply content includes:

[0111] Embedding the at least one initial reply content to obtain a word embedding vector and an overall text vector of the at least one initial reply content;

[0112] Inputting the word embedding vector into an encoder, encoding it by the encoder, and obtaining an encoded latent vector corresponding to the word embedding vector;

[0113] Inputting the whole text vector into the memory network to obtain an output latent vector corresponding to the whole text vector;

[0114] Inputting the encoded latent vector and the output latent vector into a decoder, and decoding them by the decoder to obtain a vocabulary probability distribution corresponding to the at least one initial reply content;

[0115] Generate target content according to the probability distribution.

[0116] In actual applications, following the above example, after obtaining the emotional tendency and category label according to the first feature extraction model, the corresponding reply text is obtained from the historical database, such as H1 to Hn, and the reply text is input into the second feature extraction model. It should be noted that the corresponding reply text obtained from the historical database can be multiple, such as from H1 to Hn. Following the above example, "Thank you for your support, we will continue to work hard" is input into the second feature extraction model. It should be noted that there is at least one reply text input into the second feature extraction model, and the second feature extraction model combines the at least one reply text with the corresponding emotional category and comment category to generate the corresponding target text. The schematic diagram of the execution steps of the second feature extraction model is shown in the figure. Figure 4 shown.

[0117] Inputting at least one reply text with a corresponding label into the second feature extraction model ensures that the second feature extraction model can accurately generate the reply text, further improving the applicability and accuracy of the generated reply text.

[0118] Step 108: Reply to the comment to be replied in the interactive page based on the target reply content.

[0119] It should be noted that the interactive page may be an application scenario for the target object. For example, if the target object is a game, the interactive page may be a customer service response page for the game.

[0120] In an optional implementation of this embodiment, replying to the comment to be replied to in the interactive page based on the reply content includes:

[0121] Determine a corresponding reply template according to the target object, and add the reply content to the reply template to form a reply text;

[0122] The reply text is displayed in the interactive page.

[0123] Using the above example, based on the input reply text "Thank you for your support, we will continue to work hard", the second feature extraction model identifies the emotional tendency and comment category carried by the reply text as "satisfaction" and "game progress", and generates the corresponding target content based on the reply text: "Thank you for your love for us!"

[0124] Automatically replying to user comments on the interactive page ensures that the target content produced can be clearly displayed to the user, ensuring practicality.

[0125] The present application embodiment provides a method for intelligent comment replying, such as Figure 4 As shown, after step 108, steps 502 to 504 may also be included.

[0126] Step 502: Submit the reply content for review, and receive the review result of the reply content.

[0127] In practical applications, the audit method can be to check the generated target content for sensitive words to exclude bad information; or it can be submitted to manual for further processing to optimize the target content.

[0128] The target content is submitted for review and manually processed to further ensure the accuracy of the generated target content and avoid the problem of inappropriate information.

[0129] Step 504: Reply to the comment to be replied in the interactive page according to the audit result, and add the audit result to the history database.

[0130] Specifically, the reviewed content is used as the target content for replying, and after the reply is completed, the target content is added to the historical database. Using the above example, after the target content "Thank you for your love for us!" is generated and reviewed, the target content is added to the historical database, which further ensures that the historical database can contain reply texts for a variety of different situations.

[0131] The present application embodiment provides a method for intelligent comment replying, such as Figure 6 As shown, steps 602-610 are included.

[0132] Step 602: Displaying a comment interaction page to the user according to the user's call request, and receiving a user comment to be replied to input by the user based on the comment interaction page.

[0133] Step 604: Input the text information in the user comment to be replied to into the first feature extraction model to obtain the emotion category and comment category in the text information.

[0134] Step 606: Filter at least one initial reply content matching the emotion category and the comment category from the history database.

[0135] Step 608: Input the at least one initial reply content into a second feature extraction model to obtain target reply content.

[0136] Step 610: Reply to the user comment to be replied in the comment interaction page based on the target reply content.

[0137] In actual applications, the intelligent comment reply method can be encapsulated to obtain an application program that can be applied to different scenarios, and corresponding reply content can be produced for the actual application scenario.

[0138] The intelligent comment reply method is encapsulated to ensure that the method can be used in a variety of different application scenarios, thereby expanding the application scope of the intelligent comment reply method.

[0139] Combination Figure 7 and Figure 8 Taking the application of an intelligent comment reply method provided by this application in the game customer service area as an example, the intelligent comment reply method is further explained. Figure 8 It is a schematic flowchart of step 702 to step 706, wherein A is the first feature extraction model. Figure 7 A processing flow chart of an intelligent comment reply method for a game customer service area provided by an embodiment of the present application is shown, which specifically includes the following steps:

[0140] Step 702: receiving user comments on the target object, and preprocessing text information contained in the user comments.

[0141] For example, combined with Figure 8 In step 802 and step 804, a user posts a user comment "I didn't get any prize after ten consecutive draws, so upset!!!" in the customer service area of ​​a game, and the user comment is pre-processed to obtain a text message of "I didn't get any prize after ten consecutive draws, so upset".

[0142] Step 704: Input the text information and a corpus corresponding to the target object into a first feature extraction model to obtain the emotion category and comment category in the text information, wherein the corpus contains at least one piece of corpus information.

[0143] Combination Figure 8 , Figure 8 A is the first feature extraction model mentioned above, including an embedding layer, an encoding layer, a voting layer and a corpus. The first feature extraction model can be implemented by a Bert model. After preprocessing the user comments, the text information and the corresponding corpus are input into the first feature extraction model. Through the embedding layer and the encoding layer, combined with the corpus, the emotion category of the user comment "I didn't get any prizes after ten consecutive draws, I feel uncomfortable" is obtained: negative, sad and the comment type: lottery.

[0144] In addition, combined Figure 8 In step 806 and step 808, in actual application, the special nouns involved in the game can be summarized from forums and other areas in advance, and a corpus of the game can be generated based on the special nouns, and the corpus can be input into the first feature extraction model. Then, a corpus corresponding to the game is obtained based on the text information, wherein the corpus includes the special nouns in the game, such as "lottery", "black water", etc. It should be noted that Figure 8 The voting layer is the re-input of the comment category and the sentiment category. The schematic diagram of the Bert model is as follows Fig. 9 As shown, 901 is a partial structural diagram of the Bert model, including N encoders and corresponding N decoders. The encoder includes a multi-head attention layer and a feedforward neural network, the multi-head attention layer is used to help the encoder pay attention to the word vector in the input mapping vector and perform vector representation, and the feedforward neural network is used to receive the output of the multi-head attention layer and once again perform vector representation on the output. The decoder includes a masked multi-head attention layer, a multi-head attention layer and a feedforward neural network, wherein the masked multi-head attention layer is used to receive the input of the training sample, and the multi-head attention layer and the feedforward neural network in the decoder are used to decode the output in the encoder and perform vector representation. The mapping vector is input into the encoder for vector representation, and the mapping vector is linearly expressed and exponentially normalized by the decoder to obtain the probability of each position vector in the mapping vector in the mapping vector. In addition, when training the Bert model, the training sample also needs to be input into the decoder. Fig. 9 902 is a schematic diagram of the association relationship of vectors in the Bert model. 902 takes the schematic phrases "The weather is nice today.", "But I'm lucky," and "I think, good" as examples, and represents the association relationship between each word vector in the schematic vector according to the connecting lines.

[0145] It should be noted that in actual applications, the emotion category can be quantified by a specific numerical value, for example, represented by a numerical value of 0-100, wherein a numerical value below 50 represents a negative emotion, a numerical value above 50 represents a positive emotion, the closer to 0, the stronger the negative emotion, and the closer to 100, the stronger the positive emotion. For example, the above-mentioned "sad" emotion can be represented by a numerical value of 33.

[0146] Step 706: re-input the emotion category and comment category into the first feature extraction model to obtain a target emotion category and a target comment category.

[0147] Combination Figure 8 , the user's emotion category and comment category are obtained according to the user's comment "I feel bad after ten consecutive draws". It should be noted that the step of re-inputting the emotion category and comment category into the first feature extraction model for iterative optimization can be performed multiple times until the value corresponding to the emotion category no longer changes, such as Figure 8 As shown in step 810, the obtained comment categories and emotion categories are input into the first feature extraction model again, which can further improve the accuracy of the corresponding values ​​of the comment categories and emotion categories.

[0148] Step 708: Filter at least one initial reply content matching the emotion category and the comment category from the history database.

[0149] According to the emotion category: sad and comment type: lottery of the user comment "I didn't win after ten consecutive draws", the corresponding reply texts B1: "Don't be discouraged, please try again" and B2: "Don't be sad, there is still a chance" are obtained from the historical database.

[0150] Step 710: Input the at least one initial reply content into a second feature extraction model to obtain target reply content.

[0151] The reply texts B1: "Don't be discouraged, please try again" and B2: "Don't be sad, there is still a chance". Input the second feature extraction model to generate the target reply content: "Don't be sad, the lottery is not over, there is still a chance".

[0152] Step 712: Submit the target reply content for review, and receive the review result of the reply content.

[0153] Submit the target reply content to the manual, and the customer service staff will review the target reply content. The customer service staff will make certain revisions to the target reply content and get the review result: "Don't be sad, the lottery is not over yet, we still have a chance!", and submit the review result to the interactive page.

[0154] Step 714: Reply to the comment to be replied in the interactive page according to the audit result, and add the audit result to the history database.

[0155] The review result "Don't be sad, the lottery is not over yet, we still have a chance!" is used to reply to the user's comment "I didn't win after ten consecutive draws, I'm so upset!!!" on the customer service reply page, and then the review result "Don't be sad, the lottery is not over yet, we still have a chance!" is saved in the historical database as a reference for generating target content next time.

[0156] like Fig.10 As shown, Fig.10 A specific application scenario flow chart of an intelligent comment reply method provided in one embodiment of the present application specifically includes the following steps.

[0157] Step 1002: The user enters a comment about Game A: "I didn't get anything out of the 10 consecutive draws. I'm so upset!!!"

[0158] Step 1004: Input the user comment: "I didn't get anything after ten consecutive draws, so sad!!!" into the first feature extraction model for processing.

[0159] Step 1006: Obtain the emotion category of the user's comment: sad and the comment category: lottery through the first feature extraction model.

[0160] Step 1008: Generate a reply content through the pointer generation network according to the emotion category of the user's comment: sad and the comment category: lucky draw: "Don't be sad, the lucky draw is not over yet, we still have a chance!"

[0161] Step 1010: In response to the user's comment "I didn't win after ten consecutive draws, so sad!!!", an automatic reply is made on the reply page: "Don't be sad, the draw is not over yet, we still have a chance!"

[0162] An intelligent comment reply method provided by the embodiment of the present application not only ensures targeted replies for different application scenarios, but also ensures that reply content can be automatically generated, thereby improving the efficiency of replying to user comments, freeing up manpower, and achieving the effect of quickly and accurately replying to user comments, thus overcoming the shortcomings of the prior art.

[0163] In addition, user comments are processed according to the neural network model, ensuring that the sentiment category and comment category of the user comments can be clearly obtained in order to make further responses.

[0164] Secondly, by submitting it for review, it is ensured that the generated target reply content does not contain bad information, which further guarantees the reliability of this method.

[0165] Again, by inputting it into the first extraction model again and identifying the numerical values ​​corresponding to the emotion categories in the user comments again through the voting layer, the accuracy and reliability of the model recognition are further guaranteed.

[0166] Finally, the target response content that has passed the review is added to the historical database, which ensures the completeness of the historical database and further improves the accuracy of the generated target response content.

[0167] Corresponding to the above method embodiment, the present application also provides an embodiment of an intelligent comment reply device, Fig.11 FIG. 1 shows a schematic diagram of the structure of a smart device provided by an embodiment of the present application. Fig.11 As shown, the device comprises:

[0168] The first receiving module 1102 is configured to receive user comments on a target object, and input text information contained in the user comments and a corpus corresponding to the target object into a first feature extraction model to obtain an emotion category and a comment category in the text information, wherein the corpus contains at least one piece of corpus information;

[0169] The first screening module 1104 is configured to screen at least one initial reply content matching the emotion category and the comment category from the history database;

[0170] The first acquisition module 1106 is configured to input the at least one initial reply content into a second feature extraction model to obtain a target reply content;

[0171] The first reply module 1108 is configured to reply to the comment to be replied in the interactive page based on the target reply content.

[0172] The first receiving module 1102 is further configured to:

[0173] The text information is preprocessed by removing duplicates and clearing garbled information.

[0174] Specifically, the first feature extraction model includes an embedding layer and an encoding layer.

[0175] The first receiving module 1102 is further configured to:

[0176] Inputting the text information and the corpus into an embedding layer in a first feature extraction model, and obtaining a text information vector of the text information and at least one corpus information vector of the corpus through the embedding layer;

[0177] Inputting the text information vector and the at least one corpus information vector into an encoding layer in a first feature extraction model, and obtaining a text information encoding vector of the text information vector and at least one corpus information encoding vector of the at least one corpus information vector through the encoding layer;

[0178] The emotion category and the comment category of the text information are obtained according to the text information encoding vector and the at least one corpus information encoding vector.

[0179] The first receiving module 1102 is further configured to:

[0180] Performing word segmentation processing on the text information vector and the at least one corpus information vector through the embedding layer to obtain word units of the text information and word units of the at least one corpus information;

[0181] Embedding the word units of the text information and the word units of the at least one piece of corpus information respectively to obtain the word vector, word vector and position vector of the text information and the word vector, word vector and position vector of the at least one piece of corpus information;

[0182] Summing the character vector, word vector and position vector of the text information to obtain a vector representation of the text information;

[0183] The character vector, word vector and position vector of the at least one piece of corpus information are summed to obtain a vector representation of the at least one piece of corpus information.

[0184] The first receiving module 1102 is further configured to:

[0185] Performing part-of-speech tagging on the text information encoding vector through the encoding layer to obtain the emotion category of the text information;

[0186] Through the encoding layer, part-of-speech tagging is performed on the text information encoding vector based on the at least one corpus information encoding vector to obtain a comment category of the text information.

[0187] Furthermore, the intelligent comment replying device further includes:

[0188] The voting subunit is configured to input text information contained in the user comments to be replied to into a first feature extraction model to obtain the emotion category and comment category in the text information, including:

[0189] Inputting the text information into an embedding layer in a first feature extraction model, and obtaining a first vector representation of the text information through the embedding layer;

[0190] Processing the first vector representation of the text information through the encoding layer to obtain a first emotion category and a first comment category of the text information;

[0191] Inputting the first emotion category into a voting layer to obtain the emotion category voting result;

[0192] Inputting the emotion category voting result and the text information into an embedding layer in a first feature extraction model, and obtaining a second vector representation of the text information through the embedding layer;

[0193] The second vector representation of the text information is processed through the encoding layer to obtain a second emotion category and a second comment category of the text information and use the second emotion category and the second comment category as the emotion category and the comment category.

[0194] Specifically, the second feature extraction model includes an encoder, a memory network and a decoder.

[0195] The first acquisition module 1106 is further configured to:

[0196] Embedding the at least one initial reply content to obtain a word embedding vector and an overall text vector of the at least one initial reply content;

[0197] Inputting the word embedding vector into an encoder, encoding it by the encoder, and obtaining an encoded latent vector corresponding to the word embedding vector;

[0198] Inputting the whole text vector into the memory network to obtain an output latent vector corresponding to the whole text vector;

[0199] Inputting the encoded latent vector and the output latent vector into a decoder, and decoding them by the decoder to obtain a vocabulary probability distribution corresponding to the at least one initial reply content;

[0200] Generate target content according to the probability distribution.

[0201] The first reply module 1108 is further configured to determine a corresponding reply template according to the target object, and add the reply content to the reply template to form a reply text;

[0202] The reply text is displayed in the interactive page.

[0203] The intelligent comment replying device further includes:

[0204] The review subunit is configured to submit the reply content for review and receive the review result of the reply content;

[0205] Reply to the comments to be replied to in the interactive page according to the audit result, and add the audit result to the historical database.

[0206] Specifically, the target object includes at least one of games, audio, commodities, articles, and videos.

[0207] An intelligent comment reply device provided by the embodiment of the present application not only ensures targeted replies for different application scenarios, but also ensures that reply content can be automatically generated, thereby improving the efficiency of replying to user comments, freeing up manpower, and achieving the effect of quickly and accurately replying to user comments, thus overcoming the shortcomings of the prior art.

[0208] In addition, user comments are processed according to the neural network model, ensuring that the sentiment category and comment category of the user comments can be clearly obtained in order to make further responses.

[0209] Secondly, by submitting it for review, it is ensured that the generated target reply content does not contain bad information, which further guarantees the reliability of this method.

[0210] Again, by inputting it into the first extraction model again and identifying the numerical values ​​corresponding to the emotion categories in the user comments again through the voting layer, the accuracy and reliability of the model recognition are further guaranteed.

[0211] Finally, the target response content that has passed the review is added to the historical database, which ensures the completeness of the historical database and further improves the accuracy of the generated target response content.

[0212] Corresponding to the above method embodiment, the present application also provides an embodiment of an intelligent comment reply device, Fig.12 FIG. 1 shows a schematic diagram of the structure of a smart device provided by an embodiment of the present application. Fig.12 As shown, the device comprises:

[0213] The second receiving module 1202 is configured to display a comment input page for the user according to a call request of the user, and receive a user comment to be replied input by the user based on the comment input page;

[0214] Input module 1204 is configured to input the text information in the user comment to be replied to into the first feature extraction model to obtain the emotion category and comment category in the text information

[0215] The second screening module 1206 is configured to screen at least one initial reply content matching the emotion category and the comment category from the history database;

[0216] The second input module 1208 is configured to input the at least one initial reply content into a second feature extraction model to obtain a target reply content;

[0217] The second reply module 1210 is configured to reply to the comment to be replied in the interactive page based on the target reply content.

[0218] Corresponding to the above method embodiment, the present application also provides a first feature extraction model training device embodiment, Fig.13 FIG. 1 shows a schematic diagram of the structure of a smart device provided by an embodiment of the present application. Fig.13 As shown, the device comprises:

[0219] The sample acquisition module 1302 is configured to acquire sample comments, emotion category labels corresponding to the sample comments, and a corpus, wherein the sample comments are comment information of a target object, and the corpus is at least one piece of corpus information corresponding to the target object;

[0220] The processing module 1304 is configured to input the sample comment and the at least one piece of corpus information into a first feature extraction model to obtain a predicted emotion category output by the first feature extraction model;

[0221] A calculation module 1306 is configured to calculate a loss value of the first feature extraction model according to the emotion category label and the predicted emotion category;

[0222] The training module 1308 is configured to train the first feature extraction model based on the loss value until a training stop condition is reached.

[0223] The training device for the first feature extraction model provided in the embodiment of the present application obtains sample comments, category labels and corpora corresponding to the sample comments, and inputs the sample comments, category labels and corpora corresponding to the sample comments into the first feature extraction model, so that the first feature extraction model can understand user comments more comprehensively and accurately; then the first feature extraction model is iteratively trained according to the predicted emotion category labels, the predicted category labels and emotion labels, and the category labels. On the basis of the comprehensive and accurate understanding of user comments, the first feature extraction model can further analyze the user comments, thereby improving the accuracy of the replies and improving the effect of model training.

[0224] The above is a schematic scheme of an intelligent comment reply device of this embodiment. It should be noted that the technical scheme of the intelligent comment reply device and the technical scheme of the intelligent comment reply method described above belong to the same concept. For details not described in detail in the technical scheme of the intelligent comment reply device, please refer to the description of the technical scheme of the intelligent comment reply method described above.

[0225] Fig.14 The structure block diagram of a computing device 1400 provided according to an embodiment of the present application is shown. The components of the computing device 1400 include but are not limited to a memory 1410 and a processor 1420. The processor 1420 is connected to the memory 1410 via a bus 1430, and the database 1450 is used to store data.

[0226] The computing device 1400 also includes an access device 1440 that enables the computing device 1400 to communicate via one or more networks 1460. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 1440 may include one or more of any type of network interface (e.g., a network interface card (NIC)) whether wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a World Wide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.

[0227] In one embodiment of the present application, the above components of the computing device 1400 and Fig.14 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Fig.14The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of the present application. Those skilled in the art may add or replace other components as needed.

[0228] Computing device 1400 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. Computing device 1400 may also be a mobile or stationary server.

[0229] The processor 1420 is used to execute the following computer executable instructions:

[0230] Receiving user comments on a target object, preprocessing text information contained in the user comments, and inputting the text information and a corpus corresponding to the target object into a first feature extraction model to obtain an emotion category and a comment category in the text information, wherein the corpus contains at least one piece of corpus information;

[0231] Filtering at least one initial reply content matching the emotion category and the comment category from a history database;

[0232] Inputting the at least one initial reply content into a second feature extraction model to obtain target reply content;

[0233] Reply to the comment to be replied in the interactive page based on the target reply content.

[0234] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above-mentioned intelligent comment reply method belong to the same concept, and the details not described in detail in the technical scheme of the computing device can be found in the description of the technical scheme of the above-mentioned intelligent comment reply method.

[0235] An embodiment of the present application further provides a computer-readable storage medium storing computer instructions, which are used when executed by a processor to:

[0236] Receiving user comments on a target object, preprocessing text information contained in the user comments, and inputting the text information and a corpus corresponding to the target object into a first feature extraction model to obtain an emotion category and a comment category in the text information, wherein the corpus contains at least one piece of corpus information;

[0237] Filtering at least one initial reply content matching the emotion category and the comment category from a history database;

[0238] Inputting the at least one initial reply content into a second feature extraction model to obtain target reply content;

[0239] Reply to the comment to be replied in the interactive page based on the target reply content.

[0240] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the above-mentioned intelligent comment reply method belong to the same concept, and the details not described in detail in the technical scheme of the storage medium can be found in the description of the technical scheme of the above-mentioned intelligent comment reply method.

[0241] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0242] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0243] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described action sequence, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0244] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0245] The preferred embodiments of the present application disclosed above are only used to help explain the present application. The optional embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the present application. The present application selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can understand and use the present application well. The present application is only limited by the claims and their full scope and equivalents.

Claims

1. An intelligent comment reply method, characterized in that: include: Receiving user comments on a target object, preprocessing text information contained in the user comments, and inputting the text information and a corpus corresponding to the target object into a first feature extraction model to obtain an emotion category and a comment category in the text information, wherein the corpus contains at least one piece of corpus information; Filtering at least one initial reply content matching the emotion category and the comment category from a history database; Inputting the at least one initial reply content into a second feature extraction model to obtain target reply content; Reply to the comments to be replied to in the interactive page based on the target reply content; Wherein, the first feature extraction model includes an embedding layer and an encoding layer; Inputting the text information contained in the user comments to be replied and the corpus corresponding to the target object into the first feature extraction model to obtain the emotion category and comment category in the text information, including: Inputting the text information and the corpus into an embedding layer in a first feature extraction model, and obtaining a text information vector of the text information and at least one corpus information vector of the corpus through the embedding layer; Inputting the text information vector and the at least one corpus information vector into an encoding layer in a first feature extraction model, and obtaining a text information encoding vector of the text information vector and at least one corpus information encoding vector of the at least one corpus information vector through the encoding layer; Obtaining the emotion category and comment category of the text information according to the text information encoding vector and the at least one corpus information encoding vector; The first feature extraction model also includes a voting layer; Inputting text information contained in the user comments to be replied to into a first feature extraction model to obtain the emotion category and comment category in the text information, including: Inputting the text information into an embedding layer in a first feature extraction model, and obtaining a first vector representation of the text information through the embedding layer; Processing the first vector representation of the text information through the encoding layer to obtain a first emotion category and a first comment category of the text information; Inputting the first emotion category into a voting layer to obtain the emotion category voting result; Inputting the emotion category voting result and the text information into an embedding layer in a first feature extraction model, and obtaining a second vector representation of the text information through the embedding layer; The second vector representation of the text information is processed through the encoding layer to obtain a second emotion category and a second comment category of the text information and use the second emotion category and the second comment category as the emotion category and the comment category.

2. The method according to claim 1, characterized in that Obtaining a text information vector of the text information and at least one corpus information vector of the corpus through the embedding layer; comprising: Performing word segmentation processing on the text information vector and the at least one corpus information vector through the embedding layer to obtain word units of the text information and word units of the at least one corpus information; Embedding the word units of the text information and the word units of the at least one piece of corpus information respectively to obtain the word vector, word vector and position vector of the text information and the word vector, word vector and position vector of the at least one piece of corpus information; Summing the character vector, word vector and position vector of the text information to obtain a vector representation of the text information; The character vector, word vector and position vector of the at least one piece of corpus information are summed to obtain a vector representation of the at least one piece of corpus information.

3. The method according to claim 1, characterized in that Obtaining the emotion category and the comment category of the text information according to the text information encoding vector and the at least one corpus information encoding vector, including: Performing part-of-speech tagging on the text information encoding vector through the encoding layer to obtain the emotion category of the text information; Through the encoding layer, part-of-speech tagging is performed on the text information encoding vector based on the at least one corpus information encoding vector to obtain a comment category of the text information.

4. The method according to any one of claims 1 to 3, characterized in that: in, The second feature extraction model includes an encoder, a memory network and a decoder; Inputting the at least one initial reply content into a second feature extraction model to obtain target reply content includes: Embedding the at least one initial reply content to obtain a word embedding vector and an overall text vector of the at least one initial reply content; Inputting the word embedding vector into an encoder, encoding it by the encoder, and obtaining an encoded latent vector corresponding to the word embedding vector; Inputting the whole text vector into the memory network to obtain an output latent vector corresponding to the whole text vector; Inputting the encoded latent vector and the output latent vector into a decoder, and decoding them by the decoder to obtain a vocabulary probability distribution corresponding to the at least one initial reply content; Generate target content according to the probability distribution.

5. The method according to any one of claims 1 to 3, characterized in that: Replying to the comment to be replied to in the interactive page based on the reply content includes: Determine a corresponding reply template according to the target object, and add the reply content to the reply template to form a reply text; The reply text is displayed in the interactive page.

6. The method according to any one of claims 1 to 3, characterized in that: After replying to the comment to be replied in the interactive page based on the reply content, the method further includes: Submitting the reply content for review, and receiving the review result of the reply content; Reply to the comments to be replied to in the interactive page according to the audit result, and add the audit result to the historical database.

7. The method according to claim 1, characterized in that The target object includes at least one of a game, an audio, a commodity, an article, and a video.

8. The method according to claim 1, characterized in that Preprocessing the text information contained in the user comments includes: The text information is preprocessed by removing duplicates and clearing garbled information.

9. The method according to claim 1, characterized in that The first feature extraction model is trained in the following way: Obtain sample comments, category labels corresponding to the sample comments, and a corpus, wherein the sample comments are comment information of a target object, the corpus is at least one piece of corpus information corresponding to the target object, and the category labels include emotion category labels and comment category labels; Inputting the sample comment and the at least one piece of corpus information into a first feature extraction model to obtain a predicted emotion category and a predicted comment category output by the first feature extraction model; Calculating a loss value of the first feature extraction model according to the emotion category label, the predicted emotion category, the comment category label, and the predicted comment category; The first feature extraction model is trained based on the loss value until a training stop condition is reached.

10. An intelligent comment reply method, characterized in that: include: Displaying a comment interaction page for the user according to the user's call request, and receiving a user comment to be replied to input by the user based on the comment interaction page; Inputting the text information in the user comments to be replied and the corpus corresponding to the target object into a first feature extraction model to obtain the emotion category and comment category in the text information; Filtering at least one initial reply content matching the emotion category and the comment category from a history database; Inputting the at least one initial reply content into a second feature extraction model to obtain target reply content; Replying to the user comments to be replied to in the comment interaction page based on the target reply content; Wherein, the first feature extraction model includes an embedding layer and an encoding layer; Inputting the text information contained in the user comments to be replied and the corpus corresponding to the target object into the first feature extraction model to obtain the emotion category and comment category in the text information, including: Inputting the text information and the corpus into an embedding layer in a first feature extraction model, and obtaining a text information vector of the text information and at least one corpus information vector of the corpus through the embedding layer; Inputting the text information vector and the at least one corpus information vector into an encoding layer in a first feature extraction model, and obtaining a text information encoding vector of the text information vector and at least one corpus information encoding vector of the at least one corpus information vector through the encoding layer; Obtaining the emotion category and comment category of the text information according to the text information encoding vector and the at least one corpus information encoding vector; The first feature extraction model also includes a voting layer; Inputting text information contained in the user comments to be replied to into a first feature extraction model to obtain the emotion category and comment category in the text information, including: Inputting the text information into an embedding layer in a first feature extraction model, and obtaining a first vector representation of the text information through the embedding layer; Processing the first vector representation of the text information through the encoding layer to obtain a first emotion category and a first comment category of the text information; Inputting the first emotion category into a voting layer to obtain the emotion category voting result; Inputting the emotion category voting result and the text information into an embedding layer in a first feature extraction model, and obtaining a second vector representation of the text information through the embedding layer; The second vector representation of the text information is processed through the encoding layer to obtain a second emotion category and a second comment category of the text information and use the second emotion category and the second comment category as the emotion category and the comment category.

11. An intelligent comment replying device, characterized in that: include: A first receiving module is configured to receive user comments on a target object, and input text information contained in the user comments and a corpus corresponding to the target object into a first feature extraction model to obtain an emotion category and a comment category in the text information, wherein the corpus contains at least one corpus information, and wherein the first feature extraction model includes an embedding layer and an encoding layer; Inputting the text information contained in the user comments to be replied and the corpus corresponding to the target object into the first feature extraction model to obtain the emotion category and comment category in the text information, including: Inputting the text information and the corpus into an embedding layer in a first feature extraction model, and obtaining a text information vector of the text information and at least one corpus information vector of the corpus through the embedding layer; Inputting the text information vector and the at least one corpus information vector into an encoding layer in a first feature extraction model, and obtaining a text information encoding vector of the text information vector and at least one corpus information encoding vector of the at least one corpus information vector through the encoding layer; Obtaining the emotion category and comment category of the text information according to the text information encoding vector and the at least one corpus information encoding vector; The first feature extraction model also includes a voting layer; Inputting text information contained in the user comments to be replied to into a first feature extraction model to obtain the emotion category and comment category in the text information, including: Inputting the text information into an embedding layer in a first feature extraction model, and obtaining a first vector representation of the text information through the embedding layer; Processing the first vector representation of the text information through the encoding layer to obtain a first emotion category and a first comment category of the text information; Inputting the first emotion category into a voting layer to obtain the emotion category voting result; Inputting the emotion category voting result and the text information into an embedding layer in a first feature extraction model, and obtaining a second vector representation of the text information through the embedding layer; Processing the second vector representation of the text information through the encoding layer to obtain a second emotion category and a second comment category of the text information and using the second emotion category and the second comment category as the emotion category and the comment category; A first screening module is configured to screen at least one initial reply content matching the emotion category and the comment category from a history database; A first acquisition module is configured to input the at least one initial reply content into a second feature extraction model to obtain a target reply content; The first reply module is configured to reply to the comment to be replied in the interactive page based on the target reply content.

12. An intelligent comment replying device, characterized in that: include: A second receiving module is configured to display a comment input page for the user according to a call request of the user, and receive a user comment to be replied input by the user based on the comment input page; An input module is configured to input the text information in the user comment to be replied and the corpus corresponding to the target object into a first feature extraction model to obtain the emotion category and comment category in the text information, wherein the first feature extraction model includes an embedding layer and an encoding layer; Inputting the text information contained in the user comments to be replied and the corpus corresponding to the target object into the first feature extraction model to obtain the emotion category and comment category in the text information, including: Inputting the text information and the corpus into an embedding layer in a first feature extraction model, and obtaining a text information vector of the text information and at least one corpus information vector of the corpus through the embedding layer; Inputting the text information vector and the at least one corpus information vector into an encoding layer in a first feature extraction model, and obtaining a text information encoding vector of the text information vector and at least one corpus information encoding vector of the at least one corpus information vector through the encoding layer; Obtaining the emotion category and comment category of the text information according to the text information encoding vector and the at least one corpus information encoding vector; The first feature extraction model also includes a voting layer; Inputting text information contained in the user comments to be replied to into a first feature extraction model to obtain the emotion category and comment category in the text information, including: Inputting the text information into an embedding layer in a first feature extraction model, and obtaining a first vector representation of the text information through the embedding layer; Processing the first vector representation of the text information through the encoding layer to obtain a first emotion category and a first comment category of the text information; Inputting the first emotion category into a voting layer to obtain the emotion category voting result; Inputting the emotion category voting result and the text information into an embedding layer in a first feature extraction model, and obtaining a second vector representation of the text information through the embedding layer; Processing the second vector representation of the text information through the encoding layer to obtain a second emotion category and a second comment category of the text information and using the second emotion category and the second comment category as the emotion category and the comment category; A second screening module is configured to screen at least one initial reply content matching the emotion category and the comment category from the history database; A second input module is configured to input the at least one initial reply content into a second feature extraction model to obtain a target reply content; The second reply module is configured to reply to the comment to be replied in the interactive page based on the target reply content.

13. A computing device comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, characterized in that: When the processor executes the instructions, the steps of the method described in any one of claims 1-9 or 10 are implemented.

14. A computer-readable storage medium, characterized in that: It stores computer instructions, which, when executed by a processor, implement the steps of the method described in any one of claims 1-9 or 10.

Images