Comment detection method and comment detection device, electronic device and storage medium

Abstract

The present application provides a comment detection method and comment detection device, electronic device and storage medium, which belong to the field of artificial intelligence technology. The method includes: obtaining the target object's comment text to be detected and the rating to be detected for the target product; performing object matching based on the target product to obtain at least one first candidate object; obtaining the candidate historical rating of each first candidate object, and performing rating comparison on each candidate historical rating to obtain at least one second candidate object; performing text similarity calculation on the comment text to be detected and the candidate comment text of each second candidate object to obtain a text similarity value; determining the target historical object from at least one second candidate object by numerically comparing the text similarity values; inputting the target behavior characteristics of the target historical object and the comment text to be detected into a pre-trained comment detection model to obtain a comment detection result. The embodiment of the present application can accurately detect the authenticity of cold-start user comments.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a comment detection method and device, an electronic device, and a storage medium. Background Art

[0002] Currently, when a product review is submitted by a new user (i.e., a cold-start user), the authenticity of the review cannot be accurately and effectively determined solely based on the cold-start user's review text due to a lack of historical behavioral information about the cold-start user. Therefore, accurately detecting the authenticity of cold-start user reviews has become a pressing technical challenge. Summary of the Invention

[0003] The main purpose of the embodiments of the present application is to propose a comment detection method and comment detection device, electronic device and storage medium, aiming to accurately detect the authenticity of cold start user comments.

[0004] To achieve the above objectives, a first aspect of an embodiment of the present application provides a comment detection method, the method comprising:

[0005] Obtain the target object's review text and rating for the target product to be tested;

[0006] Performing object matching on the target product based on a pre-built historical review library to obtain at least one historical review object that has reviewed the target product, and using the historical review object as a first candidate object; the historical review library is used to store all products to be reviewed, review information of the products to be reviewed, and review objects of the review information;

[0007] Obtaining a candidate historical rating for each of the first candidate objects, performing a rating comparison on each of the candidate historical ratings based on the rating to be detected and a preset comparison condition, to obtain at least one second candidate object, where the second candidate object is used to represent the first candidate object for which a result of the rating comparison satisfies the preset comparison condition;

[0008] Obtaining a candidate comment text for each second candidate object, performing text similarity calculation on the comment text to be detected and the candidate comment text to obtain a text similarity value;

[0009] By performing a numerical comparison on the text similarity values, determining a target historical object from at least one of the second candidate objects according to a result of the numerical comparison;

[0010] Obtaining target behavior characteristics of the target history object;

[0011] The target behavior feature and the comment text to be detected are input into a pre-trained comment detection model to obtain a comment detection result.

[0012] In some embodiments, the preset comparison condition includes a preset rating threshold, and obtaining the candidate historical rating of each first candidate object, performing a rating comparison on each candidate historical rating based on the rating to be detected and the preset comparison condition to obtain at least one second candidate object includes:

[0013] Calculating the difference between the candidate historical rating and the rating to be tested to obtain a candidate rating difference;

[0014] Performing numerical comparison on the candidate rating differences according to a preset comparison condition to obtain a numerical comparison result;

[0015] When the numerical comparison result indicates that the candidate rating difference is less than or equal to the preset rating threshold, the first candidate object corresponding to the numerical comparison result is used as the second candidate object.

[0016] In some embodiments, calculating text similarity between the to-be-detected comment text and the candidate comment text to obtain a text similarity value includes:

[0017] Performing text segmentation processing on the to-be-detected comment text and the candidate comment text respectively to obtain a first word text corresponding to the to-be-detected comment text and a second word text corresponding to the candidate comment text;

[0018] Encoding the first word text and the second word text respectively to obtain a first text word vector corresponding to the first word text and a second text word vector corresponding to the second word text object;

[0019] A text similarity calculation is performed on the first text word vector and the second text word vector to obtain a text similarity value.

[0020] In some embodiments, the target behavior feature includes a quantity behavior feature, a length behavior feature, and a comment behavior feature, and obtaining the target behavior feature of the target history object includes:

[0021] Performing a first comment processing on the target historical object to determine the quantity behavior feature;

[0022] performing a second comment processing on the target history object to determine the length behavior feature;

[0023] Performing a third comment processing on the target historical object to determine the comment behavior characteristics;

[0024] The quantity behavior feature, the length behavior feature, and the comment behavior feature are merged to obtain a target behavior feature.

[0025] In some embodiments, performing the first comment processing on the target historical object to determine the quantity behavior feature includes:

[0026] Counting the number of comments on each historical comment object within a preset unit period to obtain a first statistical result, where the first statistical result includes the total number of historical comments on each historical comment object within the preset unit period;

[0027] Comparing the total number of historical reviews to determine a peak number of historical reviews;

[0028] Determine a target peak number of comments on the target historical object in a preset unit period according to the first statistical result;

[0029] The quantity behavior feature is determined according to the peak value of the number of historical comments and the peak value of the number of target comments.

[0030] In some embodiments, performing the second comment processing on the target history object to determine the length behavior feature includes:

[0031] Acquire at least one target comment text of the target history object, perform text length statistics on the target comment text, and obtain a second statistical result, wherein the second statistical result includes a comment length value of each target comment text;

[0032] Comparing the lengths of the comments to determine a historical length peak;

[0033] Performing historical comment matching on the target historical object to obtain at least one historical comment text;

[0034] The length behavior feature is determined according to the historical comment text and the historical length peak value.

[0035] In some embodiments, performing the third comment processing on the target historical object to determine the comment behavior characteristics includes:

[0036] Acquire at least one target history evaluation information of the target history object;

[0037] Comparing the target historical evaluation information according to a preset comment judgment threshold to obtain a rating comparison result;

[0038] The review behavior characteristics are determined according to the rating comparison results.

[0039] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application provides a comment detection device, comprising:

[0040] The target object acquisition module is used to obtain the target object's review text and rating to be tested on the target product;

[0041] An object matching module is configured to perform object matching on the target product based on a pre-built historical review library, obtain at least one historical review object that has reviewed the target product, and use the historical review object as a first candidate object; the historical review library is configured to store all products to be reviewed, review information of the products to be reviewed, and review objects of the review information;

[0042] a rating comparison module, configured to obtain a candidate historical rating for each of the first candidate objects, perform a rating comparison on each of the candidate historical ratings based on the rating to be detected and a preset comparison condition, and obtain at least one second candidate object, where the second candidate object is used to represent the first candidate object for which the rating comparison result satisfies the preset comparison condition;

[0043] a text similarity calculation module, configured to obtain a candidate comment text for each second candidate object, perform text similarity calculation on the comment text to be detected and the candidate comment text, and obtain a text similarity value;

[0044] a numerical comparison module, configured to perform numerical comparison on the text similarity values ​​and determine a target historical object from at least one of the second candidate objects according to a result of the numerical comparison;

[0045] A behavior feature acquisition module, configured to acquire target behavior features of the target historical object;

[0046] The comment detection module is used to input the target behavior characteristics and the comment text to be detected into a pre-trained comment detection model to obtain a comment detection result.

[0047] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the method described in the first aspect when executing the computer program.

[0048] To achieve the above-mentioned purpose, the fourth aspect of the embodiments of the present application proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the method described in the first aspect.

[0049] This application proposes a review detection method, review detection device, electronic device, and storage medium. The method obtains a target object's review text and rating to be detected for a target product, performs object matching on the target product based on a pre-built historical review library, obtains at least one historical review object that has reviewed the target product, and uses this historical review object as a first candidate object. The historical review library is used to store all products to be reviewed, their review information, and the review objects of the review information, to identify historical review objects that have reviewed the target product. A candidate historical rating is obtained for each first candidate object, and a rating comparison is performed on each candidate historical rating based on the rating to be detected and a preset comparison condition to obtain at least one second candidate object. The second candidate object represents a first candidate object whose rating comparison result satisfies the preset comparison condition. Then, a candidate review text is obtained for each second candidate object, and a text similarity calculation is performed between the review text to be detected and the candidate review text to obtain a text similarity value. A numerical comparison of the text similarity values ​​is performed, and a target historical object is determined from the at least one second candidate object based on the result of the numerical comparison. The target historical object represents the historical review object with the most similar review status to the target object. Finally, the target behavior features of the target historical object are obtained, and the target behavior features and the comment text to be detected are input into a pre-trained comment detection model to obtain the comment detection result. The embodiment of the present application can improve the accuracy of the authenticity detection of cold start user comments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of the review detection method provided in the embodiment of the present application;

[0051] Figure 2 yes Figure 1 Flow chart of step S130 in FIG.

[0052] Figure 3 yes Figure 1 Flow chart of step S140 in FIG.

[0053] Figure 4 yes Figure 1 Flow chart of step S160 in FIG.

[0054] Figure 5 yes Figure 4 Flow chart of step S410 in FIG.

[0055] Figure 6 yes Figure 4 Flow chart of step S420 in FIG.

[0056] Figure 7 yes Figure 4 Flow chart of step S430 in FIG.

[0057] Figure 8 This is a schematic diagram of the structure of the behavior characteristics of updating the target object provided by the embodiment of the present application;

[0058] Figure 9 Schematic diagram of the structure of the comment detection device provided in the embodiment of the present application;

[0059] Figure 10 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0061] It should be noted that although the device schematics illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the device or the sequence in the flowcharts. The terms "first," "second," and so on, in the specification, claims, and drawings, are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0063] First, let’s analyze some of the terms used in this application:

[0064] Artificial Intelligence (AI) is a new technical discipline that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. A branch of computer science, AI seeks to understand the essence of intelligence and create new intelligent machines that can respond in a manner similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thinking. It also encompasses theories, methods, technologies, and application systems that use digital computers or digital computer-controlled machines to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results.

[0065] BERT (Bidirectional Encoder Representation from Transformers): It is used to further enhance the generalization capabilities of word embedding models, fully describing features at the character, word, sentence, and even inter-sentence levels. It is built on the Transformer. BERT has three types of embeddings: Token Embedding, Segment Embedding, and Position Embedding. Token Embeddings are word embeddings, with the first word being the CLS marker, which can be used for subsequent classification tasks. Segment Embeddings are used to distinguish between two sentences, as pre-training involves not only LM but also classification tasks with two sentences as input. Position Embeddings, where the positional word vectors are not the trigonometric functions used in Transformers, but are learned by BERT through training. However, BERT directly trains a Position Embedding to preserve position information. A vector is randomly initialized for each position and added to the model training, resulting in an embedding containing position information. BERT then directly concatenates these Position Embeddings with the Word Embeddings.

[0066] Cosine distance: also known as cosine similarity, uses the cosine value of the angle between two vectors in the vector space as a measure of the size of the difference between two individuals.

[0067] With the rapid development of internet technology and the e-commerce industry, people are increasingly reliant on online purchases. This means buyers are unable to physically touch the products, making it difficult to assess factors influencing their purchase, such as their quality. Consequently, product reviews have become a crucial tool for assessing product quality. However, with the increasing number of fake reviews, the detection of fake reviews is hampered by the inability to effectively identify the authenticity of a review from a new user, often a cold-start user, due to a lack of historical behavioral data. Therefore, accurately detecting the authenticity of cold-start user reviews has become a pressing technical challenge.

[0068] Based on this, the embodiments of the present application provide a comment detection method and comment detection device, an electronic device and a storage medium, which aim to accurately detect the authenticity of cold start user comments.

[0069] The comment detection method, comment detection device, electronic device, and storage medium provided in the embodiments of the present application are specifically described through the following embodiments. First, the comment detection method in the embodiments of the present application is described.

[0070] The comment detection method provided in the embodiment of the present application can be applied to a terminal, can be applied to a server side, and can also be software running in a terminal or a server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or as a server cluster or distributed system composed of multiple physical servers, or as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms; the software can be an application that implements the comment detection method, etc., but is not limited to the above forms.

[0071] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0072] It should be noted that in each specific embodiment of the present application, when it comes to the need to perform relevant processing based on data related to the user's identity or characteristics, such as user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first, and the collection, use, and processing of such data will comply with relevant laws, regulations, and standards. In addition, when the embodiment of the present application needs to obtain the user's sensitive personal information, the user's separate permission or consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the user's separate permission or consent, the necessary user-related data for the normal operation of the embodiment of the present application will be obtained.

[0073] See also Figure 1 , Figure 1 This is an optional flowchart of the comment detection method provided in the embodiment of the present application. Figure 1The method may include but is not limited to steps S110 to S170:

[0074] Step S110, obtaining the target object's review text and rating to be tested on the target product;

[0075] Step S120, performing object matching on the target product according to the pre-built historical review database, obtaining at least one historical review object that has reviewed the target product, and using the historical review object as the first candidate object;

[0076] Step S130, obtaining a candidate historical rating of each first candidate object, performing a rating comparison on each candidate historical rating according to the rating to be detected and a preset comparison condition, and obtaining at least one second candidate object;

[0077] Step S140, obtaining candidate comment texts for each second candidate object, performing text similarity calculation on the comment text to be detected and the candidate comment texts, and obtaining a text similarity value;

[0078] Step S150 , determining a target historical object from at least one second candidate object based on a result of numerical comparison of the text similarity values;

[0079] Step S160, obtaining target behavior features of the target historical object;

[0080] Step S170 , inputting the target behavior features and the comment text to be detected into a pre-trained comment detection model to obtain a comment detection result.

[0081] It should be noted that the application scenarios of the embodiments of the present application may include user-end devices and server-end devices. The user-end devices are used to send the comment text to be detected and the rating to be detected of the target product to the server-end device. The server-end device is used to execute the comment detection method provided by the embodiments of the present application. After obtaining the comment text to be detected and the rating to be detected sent by the user-end device, the comment detection method of the present application is executed.

[0082] In step S110 of some embodiments, if the server-side device is used as the executor of the comment detection method provided in the embodiments of the present application, first, a communication connection is established between the server and the user side, and the target object's comment text to be detected and the rating to be detected on the target product on the user-side device are obtained based on the communication connection. Among them, the comment text to be detected is used to represent the comment text of the target object after purchasing the target product, and the rating to be detected is used to represent the score of the target object's satisfaction with the target product. The rating to be detected can be in the form of a rating or a scoring form. For example, when the satisfaction level of the product is scored in the form of a rating, the preset rating range can be 1 to 5, and the rating to be detected can be any integer score from 1 to 5; when the satisfaction level of the product is scored in the form of a score, the preset rating range can be 0.0 to 5.0, and the rating to be detected can be any score from 0.0 to 5.0.

[0083] It should be noted that the preset rating range in the form of rating is not limited to 1, 2, 3, 4, and 5. More or fewer rating levels can be set according to actual needs. The preset rating range in the form of scoring is not limited to 0.0 to 5.0. More or fewer rating levels can be set according to actual needs.

[0084] It should be noted that the embodiments of the present application are applicable to the comment detection method in all e-commerce stores and independent purchasing websites in the e-commerce platform, wherein the comment text to be detected and the rating to be detected of the target object can be expressed as the first comment of the target object after the first purchase of the target product, or can also be expressed as the first comment of the target object after a non-first purchase of the target product.

[0085] In step S120 of some embodiments, since the target object, i.e., the cold-start user, has only generated one comment behavior, it is impossible to determine the behavioral characteristics that can fully represent the user. Therefore, the problem of lack of cold-start user behavioral characteristics can be solved by determining the non-cold-start user with the highest similarity to the target object. First, a pre-constructed historical review library is obtained, which stores all the products to be reviewed, all the review information of each product to be reviewed, and all the review objects corresponding to the review information. The target product is matched according to the pre-constructed historical review library, that is, at least one historical review object that has commented on the target product is matched, and the historical review object is used as the first candidate object. Specifically, by obtaining the product number corresponding to the target product, matching is performed in the historical review library according to the product number to obtain at least one historical review object, and the historical review object is the non-cold-start user corresponding to the target product. The embodiment of the present application can perform a first screening of multiple historical review objects corresponding to the target product according to the product number, and can better screen out non-cold-start users with higher similarity, that is, at least one first candidate object.

[0086] It should be noted that the historical review database stores review information for multiple pending products, with each pending product assigned a corresponding product number to facilitate matching queries. The review information for each pending product includes at least one review subject that reviewed the pending product, the historical review text corresponding to the review subject, and a historical review score. The historical review score can be in the form of a rating or a score. The setting of the historical review score is the same as that of the pending rating described above and will not be repeated here.

[0087] In step S130 of some embodiments, since the conclusions obtained through statistics and analysis of massive data and past research work all indicate that there is a strong correlation between the score / rating of the product to be reviewed and the quality of the product / service, the candidate historical rating of each first candidate object is obtained. Afterwards, the rating comparison of each candidate historical rating is performed based on the rating to be detected and the preset comparison conditions to obtain at least one second candidate object. The embodiment of the present application performs a second screening of multiple historical review objects corresponding to the target product based on the rating comparison of the rating to be detected and the preset comparison conditions, which can further screen out non-cold start users with higher similarity, that is, at least one second candidate object.

[0088] It should be noted that the second candidate object is used to represent the first candidate object whose rating comparison result meets the preset comparison condition.

[0089] It should be noted that the candidate historical ratings can also be in the form of ratings or scores. The settings of the candidate historical ratings are the same as the settings of the above-mentioned ratings to be tested, and will not be repeated here.

[0090] See also Figure 2 , Figure 2 This is an optional flowchart of step S130 provided in an embodiment of the present application. In some embodiments, the preset comparison condition includes a preset rating threshold, and step S130 may specifically include but is not limited to steps S210 to S230:

[0091] Step S210, calculating the difference between the candidate's historical rating and the rating to be tested to obtain the candidate's rating difference;

[0092] Step S220, performing numerical comparison on the candidate rating differences according to a preset comparison condition to obtain a numerical comparison result;

[0093] Step S230: When the numerical comparison result indicates that the candidate rating difference is less than or equal to the preset rating threshold, the first candidate object corresponding to the numerical comparison result is used as the second candidate object.

[0094] In step S210 of some embodiments, when the target product is product p, the target object c's rating to be detected for product p is obtained, and the rating to be detected is recorded as M p,c Obtain the candidate historical rating of each first candidate object, calculate the difference between the candidate historical rating and the rating to be detected, and obtain the candidate rating difference.

[0095] It can be understood that, for example, if there are K first candidate objects, K is an integer greater than or equal to 1, and i∈[1,K], the candidate history rating M of the first candidate object i can be obtained. p,i , then the difference between the candidate historical rating and the rating to be tested is calculated to obtain the candidate rating difference, which can be recorded as |M p,c -M p,i |, the candidate rating difference can take the absolute value, or only retain the numerical part according to a preset function.

[0096] In some embodiments, in steps S220 to S230, the preset comparison condition includes a preset rating threshold, and the candidate rating differences are numerically compared according to the preset rating threshold of the preset comparison condition to obtain a numerical comparison result. For example, assuming that the preset rating threshold is a value of 1, when the numerical comparison result indicates that the candidate rating difference |M p,c -M p,i | is less than or equal to the preset rating threshold, that is, the target object c's rating M to be detected p,c The candidate history rating M of the first candidate object i p,i The difference is less than or equal to 1, it can be considered that the comment behavior of the target object c and the first candidate object i has a certain similarity, and the first candidate object i can be used as the second candidate object; and when the numerical comparison result indicates that the candidate rating difference |M p,c -M p,i | is greater than the preset rating threshold, that is, the target object c's rating to be detected M p,c The candidate history rating M of the first candidate object i p,i If the difference is greater than 1, it can be considered that the comment behavior of the target object c has no similarity with that of the first candidate object i, and the first candidate object i is excluded. The embodiment of the present application performs a second screening of multiple historical comment objects corresponding to the target product based on the comparison of the rating to be detected and the preset comparison condition, and can further screen out a second candidate object with higher similarity.

[0097] It should be noted that, specifically, the specific process of numerically comparing the candidate rating differences according to the preset comparison conditions is shown in formula (1), where S is used to represent a rating comparison coefficient, and S includes 0 and 1. When the numerical comparison result represents the candidate rating difference |M p,c -M p,i| is less than or equal to the preset rating threshold, the evaluation comparison coefficient S corresponding to the first candidate object is 1; when the numerical comparison result indicates that the candidate rating difference |M p,c -M p,i | is greater than a preset rating threshold, the evaluation comparison coefficient S corresponding to the first candidate object is 0. Then, the first candidate object with an evaluation comparison coefficient S of 1 is screened out as the second candidate object. By setting preset comparison conditions and evaluation comparison coefficients, the present application can improve the screening speed of the first candidate object while screening out more similar second candidate objects, thereby improving the accuracy of detecting cold start user comments.

[0098]

[0099] It should be noted that the preset rating threshold is not limited to the difference between the candidate historical rating and the to-be-detected rating of each first candidate object being less than or equal to 1, and the preset rating threshold can be flexibly adjusted according to actual needs.

[0100] In step S140 of some embodiments, in the absence of user history information, the most effective comparison is to analyze the similarity of the comment content. To further narrow the scope and reduce the amount of computation, the candidate comment text of each second candidate object is obtained, and a text similarity calculation is performed on the comment text to be tested and the candidate comment text to obtain a text similarity value.

[0101] See also Figure 3 , Figure 3 This is an optional flowchart of step S140 provided in an embodiment of the present application. In some embodiments, step S140 may specifically include but is not limited to steps S310 to S330:

[0102] Step S310, performing text segmentation processing on the review text to be detected and the candidate review texts respectively to obtain a first word text corresponding to the review text to be detected and a second word text corresponding to the candidate review text;

[0103] Step S320: Encode the first word text and the second word text respectively to obtain a first text word vector corresponding to the first word text and a second text word vector corresponding to the second word text object;

[0104] Step S330 , performing text similarity calculation on the first text word vector and the second text word vector to obtain a text similarity value.

[0105] In step S310 and step S320 of some embodiments, in order to improve the speed of text similarity calculation, the comment text to be detected and the candidate comment text are respectively input into a pre-trained text encoding model, and the text encoding model includes a segmentation layer and an encoding layer. Specifically, the comment text to be detected and the candidate comment text are respectively subjected to text segmentation processing by the segmentation layer of the text encoding model to obtain a first word text corresponding to the comment text to be detected and a second word text corresponding to the candidate comment text. Then, the first word text and the second word text are respectively encoded by the encoding layer of the text encoding model to obtain a first text word vector corresponding to the first word text and a second text word vector corresponding to the second word text object. For example, the first text word vector can be denoted as X, X=(X1, X1,…, X n ), the second text word vector can be recorded as Y, Y=(Y1,Y1,…,Y n ), where n represents the vector dimension.

[0106] It should be noted that the text encoding model can use models such as the BERT model, the language embedding model (ELMo), the generative pre-training (GPT) model to process the input comment text, and no specific limitation is made here.

[0107] In step S330 of some embodiments, a text similarity calculation is performed on the obtained first text word vector and the second text word vector to obtain a text similarity value. Specifically, the obtained word vector matrix can be analyzed using a cosine distance to determine the similarity between the to-be-detected comment text and the candidate comment text. That is, the cosine similarity of the first text word vector X and the second text word vector Y is calculated, as shown in formula (2), to obtain a text similarity value cosθ, which is used to represent the similarity between the to-be-detected comment text and the candidate comment text.

[0108]

[0109] It should be noted that, in addition to cosine distance for similarity calculation, the embodiment of the present application may also adopt calculation methods such as Euclidean distance and Pearson correlation coefficient, which are not specifically limited here.

[0110] In step S150 of some embodiments, after performing text similarity calculation on the comment text to be detected and the candidate comment text of each second candidate object, and obtaining the text similarity value corresponding to each second candidate object, a numerical comparison is performed on the obtained text similarity values, and the second candidate object with the largest text similarity value is determined based on the result of the numerical comparison, that is, the target historical object is determined, and the comment information corresponding to the target historical object is the historical comment object that is most similar to the comment information of the target object. By performing text similarity calculation on the candidate comment text of the second candidate object and the comment text to be detected, the embodiment of the present application can further screen out the historical comment object that is most similar to the target object from the second candidate objects, thereby accurately detecting the authenticity of the cold start user comments.

[0111] In step S160 of some embodiments, the present application may simultaneously consider explicit external behavior information and implicit text feature information. Specifically, by matching the target product, comparing the ratings of the first candidate object with the target object, and comparing the text similarity values ​​of the second candidate object with the target object, the target historical object with the highest similarity is obtained. The target behavior features of the target historical object are then obtained to update the cold start user's features based on the target behavior features, thereby resolving the problem of a lack of behavioral features for cold start users.

[0112] See also Figure 4 , Figure 4 This is an optional flow chart of step S160 provided in an embodiment of the present application. In some embodiments, the target behavior features include quantity behavior features, length behavior features, and comment behavior features, and step S160 may specifically include but is not limited to steps S410 to S440:

[0113] Step S410 , performing a first comment processing on the target historical object to determine the quantity behavior characteristics;

[0114] Step S420 , performing a second comment processing on the target history object to determine the length behavior feature;

[0115] Step S430, performing a third comment processing on the target historical object to determine comment behavior characteristics;

[0116] Step S440 , merging the quantity behavior feature, the length behavior feature, and the comment behavior feature to obtain the target behavior feature.

[0117] In some embodiments, in steps S410 to S440, in order to enable the cold-start user, i.e., the target object, to also have more complete behavioral characteristics to achieve subsequent effective sorting and other judgments, it is necessary to fully obtain the target behavioral characteristics of the target historical object that is most similar to the target object. The target behavioral characteristics include quantity behavioral characteristics, length behavioral characteristics, and comment behavioral characteristics, wherein the quantity behavioral characteristics are used to characterize the behavioral characteristics of the quantity information of the target historical object in the process of historical comments on the product; the length behavioral characteristics are used to characterize the behavioral characteristics of the comment length information of the target historical object in the process of historical comments on the product; and the comment behavioral characteristics are used to characterize the behavioral characteristics of the rating / score information of the target historical object in the process of historical comments on the product.

[0118] It should be noted that the target behavior feature of the target history object can be obtained based on the analysis of all historical reviews of different commodities, or can be obtained based on the analysis of all historical reviews of the target commodity only.

[0119] See also Figure 5 , Figure 5 This is an optional flowchart of step S410 provided in an embodiment of the present application. In some embodiments, step S410 may specifically include but is not limited to steps S510 to S540:

[0120] Step S510: Counting the number of comments on each historical comment object within a preset unit period to obtain a first statistical result, where the first statistical result includes the total number of historical comments on each historical comment object within the preset unit period;

[0121] Step S520, comparing the total number of historical comments to determine the peak number of historical comments;

[0122] Step S530, determining a target comment quantity peak value of the target historical object in a preset unit period according to the first statistical result;

[0123] Step S540: determining quantity behavior characteristics based on the peak number of historical comments and the peak number of target comments.

[0124] In step S510 of some embodiments, a pre-built historical review library is obtained, and a counter is set for each historical review object. Then, the number of comments on each historical review object within a preset unit period can be counted based on the historical review library to obtain a first statistical result, which includes the total number of historical comments on each historical review object in the preset unit period.

[0125] It should be noted that the historical review library will record the historical reviews of all historical review objects in a preset unit period.

[0126] It should be noted that the preset unit period can be one day, that is, the sum of the number of comments published by each historical comment object in a single day is counted as the total number of historical comments. In addition, the preset unit period can also be one week, which can be adjusted according to actual conditions and is not specifically limited here.

[0127] In some embodiments, in steps S520 and S530, the total number of historical reviews for all historical review objects is numerically compared to determine the maximum total number of historical reviews, which is the peak number of historical reviews. A target peak number of reviews for the target historical object in a preset unit period is then determined based on the first statistical result. The target peak number of reviews represents the maximum number of reviews for the target historical object in the preset unit period.

[0128] In step S540 of some embodiments, after determining the peak number of historical comments and the peak number of target comments, as shown in formula (3), the quantity behavior feature is determined based on the peak number of historical comments and the peak number of target comments, that is, the target peak number of comments of the target historical object in a preset unit period is compared with the peak number of historical comments to obtain the quantity behavior feature, which can be denoted as F.

[0129]

[0130] Among them, r represents the number of comments, Max(r a ) is used to represent the maximum number of comments on the target historical object a in a preset unit period, that is, the peak number of target comments, and DB is used to represent the historical comment database. Used to indicate the maximum number of comments made by a commenting user in the historical comment library in a preset unit period, that is, the peak number of historical comments.

[0131] It should be noted that since the number of comments posted by a comment user within a preset unit period can reflect his abnormal behavior, for example, the number of comments posted by a normal comment user in a day mostly does not exceed 5, while an abnormal comment user needs to post a large number of false comments to achieve his goal, the embodiment of the present application can determine the abnormality of the number of comments of the target historical object based on the obtained quantity behavior feature F.

[0132] See also Figure 6 , Figure 6 This is an optional flowchart of step S420 provided in an embodiment of the present application. In some embodiments, step S420 may specifically include but is not limited to steps S610 to S640:

[0133] Step S610: obtaining at least one target comment text of the target history object, performing text length statistics on the target comment text, and obtaining a second statistical result, the second statistical result including the comment length value of each target comment text;

[0134] Step S620, comparing the comment length values ​​to determine the historical length peak;

[0135] Step S630, performing historical comment matching on the target historical object to obtain at least one historical comment text;

[0136] Step S640: determining length behavior features based on historical comment texts and historical length peaks.

[0137] In some embodiments, in steps S610 and S620, at least one target comment text of a target historical object is obtained from a pre-built historical comment library, and text length statistics are performed on each target comment text to obtain a second statistical result, which includes the comment length value of each target comment text. Subsequently, the obtained comment length values ​​are numerically compared to determine the maximum comment length value in the historical comment library, which is the historical length peak.

[0138] In steps S630 to S640 of some embodiments, historical comments of the target historical object are matched according to the historical comment library to obtain at least one historical comment text. Text length statistics are performed on the at least one historical comment text obtained, and the target comment length peak, target comment length valley and target comment length mean of the target historical object are determined based on the statistical results. Among them, the target comment length peak is used to represent the longest comment length in the historical comment text of the target historical object; the target comment length valley is used to represent the shortest comment length in the historical comment text of the target historical object; and the target comment length mean is used to represent the average comment length in the historical comment text of the target historical object. Finally, as shown in formula (4), the length behavior feature is determined based on the historical comment text and the historical length peak, that is, the length behavior feature L is determined based on the target comment length peak, the target comment length valley, the target comment length mean and the historical length peak.

[0139]

[0140] Among them, len1(r a ) is used to represent the target comment length peak of the target history object, len2(r a ) is used to represent the target comment length valley of the target history object, len3(r a ) is used to represent the target comment length mean of the target history object, It is used to represent the maximum comment length of comment users in the historical comment library, that is, the historical length peak value. α, β, and γ respectively represent the weighting coefficients of the corresponding ratios, which can be set to 1 / 3 on average or adjusted according to actual needs. No specific limitation is given here.

[0141] It should be noted that, since the length of a commenting user's comment can also reflect the user's subjective behavior information, the embodiment of the present application can determine the abnormality of the comment length of the target historical object based on the obtained quantitative behavior feature L.

[0142] See also Figure 7 , Figure 7 This is an optional flowchart of step S430 provided in an embodiment of the present application. In some embodiments, step S430 may specifically include but is not limited to steps S710 to S730:

[0143] Step S710, obtaining at least one target history evaluation information of a target history object;

[0144] Step S720, comparing the target historical evaluation information according to a preset comment judgment threshold to obtain a rating comparison result;

[0145] Step S730: Determine the review behavior characteristics based on the rating comparison result.

[0146] In some embodiments, in steps S710 to S730, at least one target historical evaluation information of a target historical object is obtained from the historical review library. This target historical evaluation information may be the target historical object's rating information or product score information. To reduce malicious solicitation, the target historical evaluation information is compared based on a preset review judgment threshold to obtain a rating comparison result, and then the review behavior feature B is determined based on the rating comparison result.

[0147] It should be noted that the preset comment judgment threshold includes a rating threshold and a score threshold. For example, when the satisfaction level of a product is scored in the form of a rating, and the preset ratings include 1, 2, 3, 4, and 5, then judging the product with a rating of 1 as the lowest rating or 5 as the highest rating is considered extreme comment behavior. The rating threshold indicates that when the target historical object's rating of the product is (2, 3, 4), it is considered non-extreme comment behavior. When the satisfaction level of a product is scored in the form of a score, and the preset ratings include 0.0 to 5.0, then judging the product with a rating of 0.0 as the lowest rating or 5.0 as the highest rating is considered extreme comment behavior. The score threshold indicates that when the target historical object's rating of the product is any value between 0.0 and 5.0, it is considered non-extreme comment behavior. As shown in formula (5), a quantitative behavior feature B of 1 indicates extreme comment behavior, and a quantitative behavior feature B of 0 indicates non-extreme comment behavior, thereby determining the quantitative behavior feature B of the target historical object.

[0148]

[0149] It should be noted that since fake reviewers may intentionally praise or denigrate a product, that is, give a rating / rating that is inconsistent with the actual situation, specifically by intentionally giving the product the highest or lowest rating when rating / rating, thereby maliciously misleading users' analysis, the present embodiment can determine the evaluation anomaly of the target historical object based on the obtained quantitative behavior feature B.

[0150] In step S170 of some embodiments, since the target object, i.e., the cold-start user, only generates a comment behavior once, it is impossible to determine the behavioral characteristics that can fully represent the user. In order to enrich the behavioral characteristics of the target object, the target behavioral characteristics can be used as the current behavioral characteristics of the target object. In order to accurately detect the authenticity of the comment text to be detected of the target object, the target behavioral characteristics and the comment text to be detected are input into a pre-trained comment detection model to obtain a comment detection result, which is used to indicate whether the comment text to be detected is a valid comment or an invalid comment.

[0151] Specifically, the comment detection model can be a binary classification model built based on the BERT algorithm and a multilayer perceptron (MLP). The target behavior features and the comment text to be detected are input into a pre-trained comment detection model. First, upon receiving the comment text to be detected, the model performs feature extraction on the text to obtain the features of the text to be detected. Next, a fused feature vector consisting of the target behavior features and the features of the text to be detected is obtained. This fused feature vector is then input into the model's input layer, where it is then passed to a hidden layer fully connected to the model's input layer to calculate the hidden value of the fused feature vector. The hidden value is normalized by the output layer in the comment detection model to output the comment detection result of the comment text to be detected. The comment detection result is the probability value corresponding to the comment text to be detected under the invalid comment label and the valid comment label respectively. By numerically comparing the two probability values, the label with a larger probability value is marked as the label corresponding to the text to be detected, that is, when the probability value under the invalid comment label is greater than the probability value under the valid comment label, the comment text to be detected is a false comment; when the probability value under the invalid comment label is less than or equal to the probability value under the invalid comment label, the comment text to be detected is a true comment, thereby determining the authenticity of the comment text to be detected.

[0152] It should be noted that multiple review texts are ranked according to the validity probability of the text to be tested to obtain the review ranking result. Multiple review texts can also be sorted in descending order by combining validity probability, sentiment characteristics, length characteristics, and publication time characteristics. The ranking results are displayed on the user end, facilitating quick access to valid review information.

[0153] It should be noted that the training process of the comment detection model is specifically to construct a training sample set for the model, which includes multiple training samples, each training sample includes sample text, sample behavior features of the sample text and the initial sample label of the sample text, and the initial sample label is any one of a valid comment label and an invalid comment label. An initial detection model is constructed based on the BERT algorithm and MLP, and each sample text and the sample behavior features of the sample text are input into the initial detection model for training. Specifically, after obtaining the sample text, the model first extracts features from the sample text to obtain sample text features. Then, the sample behavior features and the sample text features are feature fused to obtain the target text features. The target text features are input into the input layer of the model, and the hidden layer fully connected to the model input layer is used to calculate the hidden value of the target text features. This hidden value is normalized by the output layer of the initial detection model to output the sample detection result for the sample text. The sample detection result is the probability value corresponding to the sample under the invalid comment label and the valid comment label, respectively. By numerically comparing the two probability values, the label with the larger probability value is marked as the label corresponding to the text to be detected, thus determining the target sample label of the sample text. The initial sample label and the target sample label are solved according to the loss function to determine the model loss value. The model parameters of the initial detection model are adjusted based on this loss value, and the adjusted initial detection model is continuously trained based on the training sample set until the loss value meets the preset training end condition to obtain the comment detection model.

[0154] It should be noted that the loss function of the model can be a 0-1 loss function, an absolute value loss function, a Hinge loss function, etc., which is not specifically limited here.

[0155] It should be noted that the preset training end condition can be when the loss value of the model is less than the preset loss value threshold, or when the similarity accuracy of the obtained initial sample label and the target sample label is greater than or equal to the preset accuracy threshold.

[0156] For example, see Figure 8, when the comment detection is performed based on the target object's comment text to be detected and the rating to be detected about the target product, that is, the target object is a cold start user, first, the comment information of the target product is obtained according to the product number of the target product, that is, at least one historical comment object that has posted comments on the same target product is obtained, recorded as the first candidate object, and the first candidate object is a non-cold start user. Then, the rating difference of the candidate historical rating and the rating to be detected of the first candidate object is screened to obtain at least one second candidate object, and the second candidate object is the first candidate object whose candidate rating difference is less than or equal to the preset rating threshold. Afterwards, according to the text similarity screening, for example, when the cosine similarity calculation is adopted, the text similarity is calculated for the comment text to be detected and the candidate comment text of each second candidate object, that is, cosθ1 to cosθ n , used to represent the corresponding text similarity value. Afterwards, the maximum value of the cosine similarity max(cosθ) is determined from the obtained text similarity value, and the second candidate object corresponding to the Max(cosθ) is the target history object. Finally, the target behavior feature of the target history object is obtained to update the behavior feature of the target object according to the target behavior feature. The embodiment of the present application designs three screening rules, that is, considering both explicit external behavior information and implicit text feature information for screening, so as to capture the non-cold start user features with the highest similarity, and embed the non-cold start user features into the cold start user, which can effectively solve the problem of lack of behavioral features of new users and new comments, thereby accurately detecting the authenticity of cold start user comments.

[0157] See also Figure 9 The present application also provides a comment detection device that can implement the above comment detection method. The device includes:

[0158] The target object acquisition module 910 is used to obtain the target object's review text and rating to be tested on the target product;

[0159] An object matching module 920 is configured to perform object matching on the target product based on a pre-built historical review database, obtain at least one historical review object that has reviewed the target product, and use the historical review object as a first candidate object; the historical review database is configured to store all products to be reviewed, review information for the products to be reviewed, and review objects of the review information;

[0160] Rating comparison module 930, configured to obtain a candidate historical rating for each first candidate object, perform a rating comparison on each candidate historical rating based on the rating to be tested and a preset comparison condition, and obtain at least one second candidate object, where the second candidate object is used to represent a first candidate object whose rating comparison result satisfies the preset comparison condition;

[0161] A text similarity calculation module 940 is used to obtain a candidate comment text for each second candidate object, perform text similarity calculation on the comment text to be detected and the candidate comment text, and obtain a text similarity value;

[0162] a numerical comparison module 950 for performing numerical comparison on text similarity values ​​and determining a target historical object from at least one second candidate object according to a result of the numerical comparison;

[0163] A behavior feature acquisition module 960 is used to acquire target behavior features of a target historical object;

[0164] The comment detection module 970 is used to input the target behavior features and the comment text to be detected into a pre-trained comment detection model to obtain the comment detection results.

[0165] The specific implementation of the comment detection device is substantially the same as the specific embodiment of the comment detection method described above, and will not be described in detail here.

[0166] The present application also provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above-mentioned comment detection method when executing the computer program. The electronic device can be any smart terminal including a tablet computer, an in-vehicle computer, or the like.

[0167] See also Figure 10 , Figure 10 The hardware structure of an electronic device according to another embodiment is shown. The electronic device includes:

[0168] The processor 1010 can be implemented as a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application.

[0169] The memory 1020 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1020 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1020 and is called by the processor 1010 to execute the comment detection method of the embodiments of this application.

[0170] Input / output interface 1030, used to implement information input and output;

[0171] Communication interface 1040, used to implement communication interaction between this device and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WiFi, Bluetooth, etc.);

[0172] bus 1050 , which transmits information between various components of the device (e.g., processor 1010 , memory 1020 , input / output interface 1030 , and communication interface 1040 );

[0173] The processor 1010 , the memory 1020 , the input / output interface 1030 , and the communication interface 1040 are connected to each other in communication within the device via a bus 1050 .

[0174] An embodiment of the present application further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned comment detection method is implemented.

[0175] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0176] The embodiments of the present application provide a review detection method and review detection device, electronic device, and storage medium. The method obtains the target object's review text to be detected and the rating to be detected for the target product, and performs object matching based on the target product in a pre-built historical review library to obtain at least one historical review object that has commented on the target product. The historical review object is used as a first candidate object to determine the historical review object that has commented on the target product, wherein the historical review library is used to store all the products to be reviewed, all the review information of each product to be reviewed, and all the review objects corresponding to the review information. The difference between the candidate historical rating and the rating to be detected is calculated to obtain a candidate rating difference. The candidate rating difference is numerically compared according to a preset comparison condition to obtain a numerical comparison result. When the numerical comparison result indicates that the candidate rating difference is less than or equal to a preset rating threshold, the first candidate object corresponding to the numerical comparison result is used as a second candidate object. The second candidate object is used to indicate that the rating comparison result satisfies the preset comparison condition. Then, the candidate review text of each second candidate object is obtained, and the text segmentation processing is performed on the review text to be detected and the candidate review text respectively to obtain the first word text corresponding to the review text to be detected and the second word text corresponding to the candidate review text. Afterwards, the first word text and the second word text are respectively encoded to obtain the first text word vector corresponding to the first word text and the second text word vector of the second word text object, and the text similarity calculation is performed on the first text word vector and the second text word vector to obtain a text similarity value. By performing a numerical comparison on the text similarity value, a target history object is determined from at least one second candidate object according to the result of the numerical comparison, and the target history object is used to characterize the historical comment object that is most similar to the comment situation of the target object. Finally, the quantity behavior characteristics, length behavior characteristics and comment behavior characteristics of the target history object are obtained, and the quantity behavior characteristics, length behavior characteristics, comment behavior characteristics and the comment text to be detected are input into a pre-trained comment detection model to obtain a comment detection result. The embodiment of the present application can mine the comment user behavior characteristics associated with the comment, that is, it simultaneously considers explicit external behavior information and implicit text feature information for screening to capture the non-cold start user features with the highest similarity, and embeds the non-cold start user features into the cold start user, which can effectively solve the problem of lack of behavior characteristics of new users and new comments, and then accurately detect the authenticity of cold start user comments.

[0177] The embodiments described in the embodiments of this application are intended to more clearly illustrate the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0178] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0179] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0180] Those skilled in the art will appreciate that all or some of the steps in the methods, systems, and functional modules / units in the devices disclosed above may be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0181] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0182] It should be understood that in this application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0183] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the above-mentioned units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0184] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0185] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0186] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes multiple instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: various media that can store programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0187] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention should be within the scope of the present invention.

Claims

1. A comment detection method, characterized in that: The method comprises: Obtain the target object's review text and rating for the target product to be tested; Performing object matching on the target product based on a pre-built historical review library to obtain at least one historical review object that has reviewed the target product, and using the historical review object as a first candidate object; the historical review library is used to store all products to be reviewed, review information of the products to be reviewed, and review objects of the review information; Obtaining a candidate historical rating for each of the first candidate objects, performing a rating comparison on each of the candidate historical ratings based on the rating to be detected and a preset comparison condition, to obtain at least one second candidate object, where the second candidate object is used to represent the first candidate object for which a result of the rating comparison satisfies the preset comparison condition; Obtaining a candidate comment text for each second candidate object, performing text similarity calculation on the comment text to be detected and the candidate comment text to obtain a text similarity value; By performing a numerical comparison on the text similarity values, determining a target historical object from at least one of the second candidate objects according to a result of the numerical comparison; Acquire target behavior features of the target history object, wherein the target behavior features include quantity behavior features, length behavior features, and comment behavior features; The target behavior features and the comment text to be detected are input into a pre-trained comment detection model to obtain a comment detection result. The comment detection model is a binary classification model constructed based on the BERT algorithm and a multi-layer perceptron.

2. The method according to claim 1, characterized in that The preset comparison condition includes a preset rating threshold, and obtaining a candidate historical rating of each first candidate object, performing a rating comparison on each candidate historical rating according to the rating to be detected and the preset comparison condition to obtain at least one second candidate object, includes: Calculating the difference between the candidate historical rating and the rating to be tested to obtain a candidate rating difference; Performing numerical comparison on the candidate rating differences according to a preset comparison condition to obtain a numerical comparison result; When the numerical comparison result indicates that the candidate rating difference is less than or equal to the preset rating threshold, the first candidate object corresponding to the numerical comparison result is used as the second candidate object.

3. The method according to claim 1, characterized in that The calculating of text similarity between the to-be-detected comment text and the candidate comment text to obtain a text similarity value includes: Performing text segmentation processing on the to-be-detected comment text and the candidate comment text respectively to obtain a first word text corresponding to the to-be-detected comment text and a second word text corresponding to the candidate comment text; Encoding the first word text and the second word text respectively to obtain a first text word vector corresponding to the first word text and a second text word vector corresponding to the second word text object; A text similarity calculation is performed on the first text word vector and the second text word vector to obtain a text similarity value.

4. The method according to any one of claims 1 to 3, characterized in that The acquiring of the target behavior feature of the target historical object includes: Performing a first comment processing on the target historical object to determine the quantity behavior feature; performing a second comment processing on the target history object to determine the length behavior feature; Performing a third comment processing on the target historical object to determine the comment behavior characteristics; The quantity behavior feature, the length behavior feature, and the comment behavior feature are merged to obtain a target behavior feature.

5. The method according to claim 4, characterized in that The performing the first comment processing on the target historical object to determine the quantity behavior feature includes: Counting the number of comments on each historical comment object within a preset unit period to obtain a first statistical result, where the first statistical result includes the total number of historical comments on each historical comment object within the preset unit period; Comparing the total number of historical reviews to determine a peak number of historical reviews; Determine a target peak number of comments on the target historical object in a preset unit period according to the first statistical result; The quantity behavior feature is determined according to the peak value of the number of historical comments and the peak value of the number of target comments.

6. The method according to claim 4, characterized in that The performing the second comment processing on the target historical object to determine the length behavior feature includes: Acquire at least one target comment text of the target history object, perform text length statistics on the target comment text, and obtain a second statistical result, wherein the second statistical result includes a comment length value of each target comment text; Comparing the lengths of the comments to determine a historical length peak; Performing historical comment matching on the target historical object to obtain at least one historical comment text; The length behavior feature is determined according to the historical comment text and the historical length peak value.

7. The method according to claim 4, characterized in that The performing the third comment processing on the target historical object to determine the comment behavior characteristics includes: Acquire at least one target history evaluation information of the target history object; Comparing the target historical evaluation information according to a preset comment judgment threshold to obtain a rating comparison result; The review behavior characteristics are determined according to the rating comparison results.

8. A comment detection device, characterized in that: The device comprises: The target object acquisition module is used to obtain the target object's review text and rating to be tested on the target product; An object matching module is configured to perform object matching on the target product based on a pre-built historical review library, obtain at least one historical review object that has reviewed the target product, and use the historical review object as a first candidate object; the historical review library is configured to store all products to be reviewed, review information of the products to be reviewed, and review objects of the review information; a rating comparison module, configured to obtain a candidate historical rating for each of the first candidate objects, perform a rating comparison on each of the candidate historical ratings based on the rating to be detected and a preset comparison condition, and obtain at least one second candidate object, where the second candidate object is used to represent the first candidate object for which the rating comparison result satisfies the preset comparison condition; a text similarity calculation module, configured to obtain a candidate comment text for each second candidate object, perform text similarity calculation on the comment text to be detected and the candidate comment text, and obtain a text similarity value; a numerical comparison module, configured to perform numerical comparison on the text similarity values ​​and determine a target historical object from at least one of the second candidate objects according to a result of the numerical comparison; A behavior feature acquisition module, configured to acquire target behavior features of the target history object, wherein the target behavior features include quantity behavior features, length behavior features, and comment behavior features; The comment detection module is used to input the target behavior features and the comment text to be detected into a pre-trained comment detection model to obtain comment detection results. The comment detection model is a binary classification model built based on the BERT algorithm and the multi-layer perceptron.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images