Information identification method, device and equipment

Abstract

The application discloses an information identification method, device and equipment, and the method comprises the following steps: obtaining comment information published by a user; identifying the comment information to obtain a first identification result of whether the comment information is information of a preset type; if the first identification result indicates that the comment information is not the information of the preset type, confirming the credibility of the user according to the comment information, and obtaining a second identification result of whether the comment information is the information of the preset type according to the credibility of the user. Through the above method, the preset type of comment information is automatically, accurately and efficiently identified.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, in particular to an information identification method, device and apparatus. BACKGROUND

[0002] With the development of the Internet, online forums and Baidu post bars are becoming increasingly popular. There are some "black users" who intentionally disrupt the order of discussion in any forum or Baidu post bar, which seriously affects the use of other normal users and reduces the quality of the Baidu post bar and forum. Moreover, so far, the "black users" have evolved from "senior black users" with reasons and grounds to "anti-string black users" who are good at disguising. The so-called "anti-string black users" refer to those who highly praise the phenomenon that they want to black (slander and insult), and even appear many extreme remarks to achieve the effect of anti-guiding public opinion.

[0003] Currently, there are mainly two ways to audit the content of the forum and Baidu post bar: (1) the first way is manual audit by administrators. The administrators find the remarks of "anti-string black users" by regularly browsing or reports from other users, define the users as "anti-string black users" and delete the remarks and prohibit the users from speaking; (2) the second way is intelligent audit by AI (artificial intelligence) model. The AI model trained judges whether the remarks of the users violate the relevant regulations, and if so, deletes the remarks.

[0004] The existing methods cannot well deal with the new form of "anti-string black users" who disrupt the forum and Baidu post bar, mainly having the following problems:

[0005] Firstly, the cost of manual audit is too high, and the efficiency is too poor. When the number of "anti-string black users" is large and the remarks are also many, the administrators cannot delete the relevant remarks and prohibit the users from speaking in time, which will lead to the rapid deterioration of the environment of the forum and Baidu post bar, and seriously affect the use experience of normal users.

[0006] Secondly, although the AI model can audit many remarks that do not comply with the regulations, it is difficult to judge the remarks of "anti-string black users". Because the remarks of "anti-string black users" are often disguised as "positive energy", the AI model cannot recognize the disguise, so that the remarks of "anti-string black users" can pass the audit and be published.

[0007] Finally, due to the disguising nature of the remarks of "anti-string black users", neither manual audit nor AI model can accurately judge whether the remarks of a person are "anti-string black" remarks, and the misjudgment rate is extremely high. If the misjudgment rate is too high to ban normal users, it will also seriously affect the use of normal users. SUMMARY

[0008] In view of the above problems, embodiments of the present application are proposed to provide an information identification method, device and equipment to overcome the above problems or at least partially solve the above problems.

[0009] According to an aspect of embodiments of the present application, there is provided an information identification method, comprising:

[0010] obtaining comment information published by a user;

[0011] identifying the comment information to obtain a first identification result of whether the comment information is information of a preset type;

[0012] if the first identification result indicates that the comment information is not the information of the preset type, confirming a user credibility according to the comment information, and obtaining a second identification result of whether the comment information is the information of the preset type according to the user credibility.

[0013] According to another aspect of embodiments of the present application, there is provided an information identification device, comprising:

[0014] an obtaining module configured to obtain comment information published by a user; a processing module configured to identify the comment information to obtain a first identification result of whether the comment information is information of a preset type; and if the first identification result indicates that the comment information is not the information of the preset type, confirm a user credibility according to the comment information, and obtain a second identification result of whether the comment information is the information of the preset type according to the user credibility.

[0015] According to still another aspect of embodiments of the present application, there is provided a computing device, comprising a processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface being capable of communicating with each other through the communication bus;

[0016] the memory is configured to store at least one executable instruction, and the executable instruction causes the processor to perform operations corresponding to the above information identification method.

[0017] According to yet another aspect of embodiments of the present application, there is provided a computer storage medium, the storage medium storing at least one executable instruction, and the executable instruction causes a processor to perform operations corresponding to the above information identification method.

[0018] According to the scheme provided by the above embodiment of the present application, by obtaining the comment information published by the user, identifying the comment information to obtain the first identification result of whether the comment information is the preset type of information, and if the first identification result indicates that the comment information is not the preset type of information, confirming the user credibility according to the comment information and the second identification result of whether the comment information is the preset type of information according to the user credibility, the problems of insufficient discrimination accuracy and large misjudgment rate in the prior art are solved, automatic, accurate and efficient identification of the "anti-blacklisting" user is realized, and the beneficial effect of making the BBS and forum into a "clean" discussion environment is achieved.

[0019] The above description is only a summary of the technical scheme of the embodiments of the present application, in order to more clearly understand the technical means of the embodiments of the present application, the content of the specification can be implemented, and in order to make the above and other purposes, characteristics and advantages of the embodiments of the present application more obvious and easy to understand, the specific implementation manner of the embodiments of the present application is described below. BRIEF DESCRIPTION OF DRAWINGS

[0020] By reading the detailed description of the preferred embodiments below, various other advantages and benefits will become clear to those skilled in the art. The drawings are only for the purpose of illustrating the preferred embodiments and are not considered as limiting the embodiments of the present application. Moreover, the same reference numerals are used to represent the same components throughout the drawings. In the drawings:

[0021] Figure 1 A flow chart of the information identification method provided by the embodiments of the present application is shown;

[0022] Figure 2 A specific comment identification flow chart provided by the embodiments of the present application is shown;

[0023] Figure 3 A comment identification flow chart provided by the embodiments of the present application is shown;

[0024] Figure 4 A language attitude identification model schematic diagram provided by the embodiments of the present application is shown;

[0025] Figure 5 A comment content identification model schematic diagram provided by the embodiments of the present application is shown;

[0026] Figure 6 Another flow chart of the information identification method provided by the embodiments of the present application is shown;

[0027] Figure 7 A comment processing algorithm model flow chart provided by the embodiments of the present application is shown;

[0028] Figure 8A specific anti-fraud user identification method provided by the embodiment of the present application is shown in a flow chart;

[0029] Figure 9 A specific anti-fraud user identification method provided by the embodiment of the present application is shown in a flow chart; Figure 8 A main module schematic diagram of a system in which the specific anti-fraud user identification method is shown is shown;

[0030] Figure 10 A structure schematic diagram of an information identification device provided by the embodiment of the present application is shown;

[0031] Figure 11 A structure schematic diagram of a computing device provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0032] Exemplary embodiments of the present application will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it is understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present application can be more thoroughly understood and so that the scope of the present application can be completely conveyed to those skilled in the art.

[0033] Figure 1 A flow chart of an information identification method provided by the embodiment of the present application is shown. As shown in the figure, Figure 1 The method includes the following steps:

[0034] Step 11, obtaining opinion information published by a user;

[0035] Step 12, identifying the opinion information to obtain a first identification result of whether the opinion information is information of a preset type;

[0036] Step 13, if the first identification result indicates that the opinion information is not information of the preset type, confirming user credibility according to the opinion information, and obtaining a second identification result of whether the opinion information is information of the preset type according to the user credibility.

[0037] The information identification method described in the embodiment, by obtaining opinion information published by a user, identifying the opinion information to obtain a first identification result of whether the opinion information is information of a preset type, and if the first identification result indicates that the opinion information is not information of the preset type, confirming user credibility according to the opinion information and obtaining a second identification result of whether the opinion information is information of the preset type according to the user credibility, can automatically, accurately and efficiently identify opinion information of anti-fraud users, and make a BBS or a forum a clean discussion environment.

[0038] In an optional embodiment of the present application, in step 12, the speech information is identified to obtain a first identification result of whether the speech information is information of a preset type, which includes:

[0039] The speech information is matched with a preset speech template. If the matching is successful, it is determined that the speech information is information of a preset type. If the matching is unsuccessful, the speech information is input into a trained preset speech discrimination model for processing, and a first identification result of whether the speech information is information of a preset type is output.

[0040] In the embodiment, matching the speech information with the preset speech template, if the matching is successful, determining that the speech information is information of a preset type, can include:

[0041] In step 121, the speech information is converted into a semantic vector by using a preset semantic conversion model.

[0042] In step 122, the semantic vector is used to query a semantic index of a pre-stored semantic vector library. If at least one speech template with a similarity to the speech information greater than a preset similarity threshold value is obtained, the matching is successful, and it is determined that the speech information is information of a preset type.

[0043] In the embodiment, after it is determined that the speech information is information of a preset type, further operations can include:

[0044] The information of the preset type is deleted or a user publishing the information of the preset type is subjected to a mute operation. The information of the preset type can be, for example, speech information of an "anti-black stringing" type.

[0045] In the embodiment, first, a semantic conversion model, a storage space in a memory of a semantic index corresponding to the semantic conversion model, and a similarity threshold value are initialized.

[0046] Second, related data is regularly sorted, and a preset speech is submitted as a speech template. The preset speech can include an "anti-black stringing" speech, but is not limited to the "anti-black stringing" speech.

[0047] Then, the submitted speech template is subjected to semantic matching by using a fine-tuned semantic conversion model. Expressions with similar semantics, even if not configured, can be matched, for example, "a certain person is the first person in the world" and "a certain person is the best in the world". A semantic vector code is obtained, and the speech information is stored in a background database. In this way, the matching accuracy and generalization ability of the speech template can be improved, and the configuration operation can be reduced, thereby saving the labor configuration cost.

[0048] In the above steps, only the speech template needs to be matched, so the actual speech template is not stored in memory, but only the semantic index is stored, which can effectively save memory space.

[0049] Then a semantic index is constructed for the historical semantic vector, a tree-shaped semantic index is constructed and stored in memory, and the semantic index can be constructed using the Annoy model, but is not limited thereto, and the storage in memory can be used for user query.

[0050] Then when the user posts a speech, speech information and the like data are submitted to the fine-tuned semantic conversion model, and the fine-tuned semantic conversion model converts the speech to be posted by the user into a semantic vector.

[0051] Finally, the semantic index is queried through the semantic vector, and N speech templates that are closest in semantics and have a similarity greater than a preset similarity threshold to the speech information are obtained.

[0052] If the similarity between the speech information and at least one speech template is greater than the threshold, it is determined that the speech information posted by the user is successfully matched with the speech template.

[0053] As shown in Figure 2 In this embodiment, a specific speech template configuration and recognition module is illustrated, which is mainly configured by an administrator in the system to configure "anti-black stringing" speech templates, such as "certain person is the king of the world", "certain person sweeps everything", and the like speech fragments; the speech templates are encoded by using a Bert model fine-tuned by training data to establish a semantic index of the semantic vector to accelerate the matching of the speech templates; when the user submits a speech, the user speech is also encoded by using the fine-tuned Bert model to obtain a semantic vector, and the semantic index is queried by using the semantic vector to check whether there is a match.

[0054] In another optional embodiment of the application, in step 12, the speech information is input into a trained preset speech discrimination model for processing, and a first recognition result of whether the speech information is preset type information is output, which can include:

[0055] In step 123, the speech information is input into a speech attitude discrimination model of a trained preset speech discrimination model for processing to obtain a first processing result; the first processing result includes a triple <speech information posted by the user, speech reply data, attitude>.

[0056] Specifically, the pre-defined speech discrimination model's attitude discrimination model is used to discriminate based on the user's posted speech and the replies to that speech, identifying the user's primary attitude towards the speech information. This is because if a user's post is "sophisticated satire"—seemingly praising but actually derogatory—it can be identified through the replies. For example, some people post praising comments like "So-and-so is a god descended to earth," while most replies are critical like "Haha, so-and-so lost again" or "so-and-so is not in good form," indicating that the user's post is actually negative. If the primary attitude is positive, the first processing result is not the pre-defined type of information; if the primary attitude is negative, the first processing result is the pre-defined type of information.

[0057] Step 124: Input the first processing result into the speech content discrimination model of the trained preset speech discrimination model for processing, and output the first identification result of whether the speech information is a preset type of information.

[0058] Furthermore, step 124 may also include:

[0059] If the first identification result indicates that the speech information is of a preset type, the preset type of information is deleted or the user who posted the preset type of information is muted; the preset type of information here may be speech information of the "anti-troll" type;

[0060] If the first identification result indicates that the speech information is not of a preset type, the first identification result is output directly or the first identification result is further processed.

[0061] like Figure 3 As shown in this embodiment, the speech discrimination module is first initialized, which mainly involves loading the two models, the preset speech discrimination model and the preset speech discrimination model, into memory.

[0062] Secondly, the aforementioned speech information is first converted into a binary tuple <user's speech information, reply>. This binary tuple is then input into the speech attitude discrimination model to identify the user's first attitude towards the speech information, resulting in a triple <user's speech information, reply, attitude>, where "attitude" is a boolean value, with 0 representing a negative attitude and 1 representing a positive attitude.

[0063] Finally, the triple <user's speech information, reply, attitude> is input into the speech content discrimination model for processing to obtain the first recognition result.

[0064] According to the first processing result, it is judged whether the speech information is information of a preset type. If the first processing result indicates that the speech information is not information of a preset type, the user speech and attitude are combined into a binary tuple <user speech information, attitude>.

[0065] In yet another optional embodiment of the application, step 124 can include:

[0066] Step 1241, converting the speech information into a binary tuple <user speech information, speech reply data>;

[0067] Step 1242, inputting the user speech information into a first adversarial network G1 of a speech attitude discrimination model of a trained preset speech discrimination model for processing to obtain a first output, the first adversarial network G1 inputting the user speech information into a first discriminator D1 to generate first target speech reply data;

[0068] Step 1243, inputting the speech reply data into a second adversarial network G2 of the speech attitude discrimination model of the trained preset speech discrimination model for processing to obtain a second output, the second adversarial network G2 inputting the speech reply data into a second discriminator D2 to generate first target speech information;

[0069] Step 1244, inputting the first output and the second output into a long short-term memory (LSTM) layer of the speech attitude discrimination model of the trained preset speech discrimination model for processing to obtain an output of the LSTM layer;

[0070] Step 1245, inputting the first target speech reply data and the first target speech information into a third discriminator D3 of the speech attitude discrimination model of the trained preset speech discrimination model, and inputting the output of the LSTM layer into the third discriminator D3 to obtain the first processing result.

[0071] Specifically, as shown in Figure 4 The speech attitude discrimination model includes adversarial networks G1 and G2, an LSTM network, discriminators D1, D2 and D3. The user speech information is input into the adversarial network G1 to generate a target speech reply, which is input into the discriminator D1 for judgment to generate a first judgment result. The speech reply is input into the adversarial network G2 to generate a target speech information, which is input into the discriminator D2. The outputs of G1 and G2 are also input into the LSTM for processing to obtain a target output result. The first judgment result, the second judgment result and the target output result are all input into the discriminator D3 for judgment processing to obtain a ternary tuple <user speech information, speech reply, attitude>.

[0072] The first attitude of the user's speech is obtained from the discriminator D3, and output 0 represents a negative attitude and 1 represents a positive attitude.

[0073] In another optional embodiment of the present application, step 125 can include:

[0074] Step 1251, input the user's speech information of the triple <user's speech information, speech reply data, attitude> into the first encoder of the speech content discrimination model of the trained preset speech discrimination model for processing to obtain a first encoding vector;

[0075] Step 1252, input the speech reply data into the second encoder of the speech content discrimination model for processing to obtain a second encoding vector;

[0076] Step 1253, input the attitude into the third encoder of the speech content discrimination model for processing to obtain a third encoding vector;

[0077] Step 1254, input the first encoding vector, the second encoding vector and the third encoding vector into the convolution layer of the speech content discrimination model for processing to obtain a convolution output;

[0078] Step 1255, input the convolution output into the enhancement layer of the speech content discrimination model for processing to obtain an enhanced output;

[0079] Step 1256, input the enhanced output into the fully connected layer of the speech content discrimination model for processing to output a first recognition result of whether the speech information is information of a preset type.

[0080] Specifically, as shown in Figure 5 The triple <user's speech information, speech reply, attitude> output by the discriminator D3 is encoded by three encoders to obtain three encoding vectors, and then a set of convolution layers CONV is used for convolution operation to fuse and extract features of the three vectors. Then an Attention layer is used to enhance the features, and finally a fully connected layer fc is used for label discrimination, and the label is: whether it is information of a preset type <yes, no>.

[0081] As shown in Figure 6 In an optional embodiment of the present application, the second recognition result of whether the speech information is information of a preset type according to the user's credibility confirmed in step 13 can include:

[0082] Step 131, obtaining relevant knowledge information from the knowledge graph according to the opinion information; in specific implementation, the relevant field BBSs in the knowledge graph can be found out, such as the BBSs in the sports field;

[0083] Step 132, determining the attitude information of the user according to the relevant knowledge information; in implementation, the opinion attitude discrimination model can be used to identify the true attitude of the user talking about different topics in different BBSs;

[0084] Step 133, determining the relationship matrix of the voters by using the voting algorithm according to the attitude information of the user; for example, the attitude of the BBSs in the relevant field can be determined according to the attitude of the opinions published by the user in the BBSs; thus, the relationship between the BBSs in the relevant field can be determined according to the attitude of the user, such as the hostile relationship between A BBS and B BBS, or the harmonious relationship between A BBS and B BBS; the relationship between multiple BBSs forms a relationship matrix;

[0085] Step 134, determining the credibility of the user according to the relationship matrix of the voters and the attitude information of the user; in specific implementation, the attitude of the opinions published by the user in the BBSs in the relevant field is voted by using the preset opinion processing algorithm, and the voting sequence of the voters is obtained, where the voters can be the BBSs, and the voting values are the attitudes of the user publishing opinions in the BBSs;

[0086] The first attitude of the user is compared with the voting sequence, the attitudes inconsistent with the first attitude of the user in the voting sequence are obtained, and an inconsistent attitude sequence N1, N2, …, Nm is obtained.

[0087] Each attitude value in the inconsistent attitude sequence is compared with a preset credibility threshold, if Ni is greater than or equal to the preset credibility threshold, the voter is determined to be a low-credibility voter, otherwise, the voter is determined to be a high-credibility voter; step 135, determining a second recognition result of whether the opinion information is information of a preset type according to the credibility of the user. In specific implementation, the voting values of the low-credibility voters are reversed to obtain a final target voting sequence; if the target voting sequence is consistent with the corresponding row in the relationship matrix, it is determined that the opinion of the user is not information of the preset type, otherwise, it is determined to be information of the preset type.

[0088] In this embodiment, after obtaining the second recognition result in step 135, the second recognition result can further include:

[0089] If the second recognition result is information of the preset type, the information of the preset type is deleted or the user publishing the opinion information of the preset type is subjected to a mute operation;

[0090] If the second recognition result is not the preset type of information, the second recognition result that is not the preset type of information is directly outputted.

[0091] In this embodiment, the first recognition result indicates that the comment information is not the preset type of information, the comment information is processed according to a preset comment processing algorithm, and a second recognition result of whether the comment information is the preset type of information is outputted, thereby further improving the accuracy of identifying whether the comment information published by the user is the preset type of information.

[0092] In a specific implementation example, the model can be trained according to the comments published by the user in multiple forums through a preset comment processing algorithm, and the comment information published by the user currently is further identified according to the model, thereby improving the accuracy of identification.

[0093] The implementation process of the comment processing algorithm is described below by taking whether a user in the sports field is a "anti-black" user as an example.

[0094] For example, a user publishes a comment in "A forum", and then the attitude of the user to the comment of A in related sports forums such as "B forum", "C forum", and the like is viewed, the related forums are taken as voters, if the attitude is positive and praising, a P vote is cast, otherwise, an N vote is cast; unlike a general voting algorithm, a voting sequence is generated in advance according to the relationship (that is, "harmonious" or "hostile") between the forums in the knowledge graph, for example, there are three voters "B forum", "C forum", and "D forum", the former two are "hostile" to "A forum", and the latter is "harmonious", therefore, the generated sequence is "NNP", when the votes cast by the voters are "NNP", it is considered that the user is not a "anti-black", otherwise, it is considered that the user is a "anti-black".

[0095] Specifically, first, whether to open an accurate judgment mode is selected according to the attitude of the first recognition result indicating that the comment information is not the preset type of information, and the opening of the accurate judgment mode can slightly improve the accuracy.

[0096] As shown in the following table, Figure 7 whether to open the accurate judgment mode is described by taking the identification of "anti-black" information in the sports field as an example.

[0097] The identification of "anti-black" information in the sports field generally only processes comments with a positive and praising attitude, because "anti-black" comments are generally disguised as "positive" comments, if the accurate judgment is opened, the comments with a negative and derogatory attitude will also be processed.

[0098] Step 1, obtaining relevant knowledge information from the knowledge graph according to the opinion information, for example, in the sports field, the user publishes opinions in "A BBS", and then finds out the relevant field BBS from the knowledge graph, such as "B BBS", "C BBS", "E BBS", "F BBS", "H BBS", "D BBS", etc., and obtains a matrix of the relationship between these "BBS", wherein 0 represents "hostility" and 1 represents "harmony", as shown in the following matrix:

[0099]

[0100] The order of the vertical axis and the horizontal axis is "A BBS", "B BBS", "C BBS", "E BBS", "F BBS", "H BBS", "D BBS", as shown in the above matrix, and the positioning of "A BBS" and "B BBS" is all 0, that is, the two BBSs are in "hostile" relationship.

[0101] Step 2, according to the obtained relevant field data, that is, multiple BBSs in the example, to determine the attitude of the user talking about "F" in "C BBS", the opinion attitude discrimination model can be used to input the user opinion data and opinion reply data in the four-tuple <C BBS ID, F, user opinion data, opinion reply data> to determine the first attitude information of the user.

[0102] Specifically, first, according to the obtained relevant field data, output a four-tuple <relevant field ID, the theme being talked about, user opinion data, opinion reply data>, and then determine the attitude of the four-tuple through the opinion attitude discrimination model.

[0103] Step 3, using the opinion attitude discrimination model to identify the true attitude of the user talking about different themes in different BBSs; for example, to get the attitude of the user talking about "F BBS" in "C BBS", the user opinion data and opinion reply data in the four-tuple <C BBS ID, F BBS, user opinion data, opinion reply data> can be input into the model to obtain the true attitude;

[0104] Step 4, we can assume that the calculation of step 3 can find "contradictions" between the first attitude of the user and the voting method of the preset opinion processing algorithm, which are counted as (N1, N2, …, Nm).

[0105] Step 5, using Ni obtained in step 4 to measure the credibility of the voter, if Ni is greater than or equal to the set credibility threshold, it is considered that the "voter" has low credibility, otherwise it is considered that the "voter" has high credibility; this is actually to determine whether a user has contradictory opinions in other BBSs, because we believe that if he is not an anti-black, his attitude should be consistent;

[0106] Step 6, according to the evaluation of the credibility of the voters, the correction of the voting is carried out, and the votes of the "voters" with low credibility are flipped over, so as to obtain the final voting sequence; if the sequence is the same as the sequence of the corresponding row in step 1, it is considered not to be "anti-black string", otherwise it is determined to be "anti-black string" and corresponding processing is carried out.

[0107] The following takes the identification of "anti-black string" information in the sports field as an example to illustrate the specific implementation of this step, for example:

[0108] Firstly, the data of all "voters" in each Baidu forum is obtained, and a four-tuple <forum ID, discussed topic, user comment data, comment reply data> is output, and the attitude of the four-tuple is determined through the comment attitude determination model, wherein the "discussed topic" refers to the "A", "C" and other Baidu forum topics.

[0109] As shown in Table 1, secondly, the relationship between two "voters" is obtained from two directions, wherein the user's comment on "E" in "C forum" in Table 1 is N, and the user's comment on "C" in "E forum" is P, which is obviously contradictory, and the value in the matrix obtained from the knowledge graph is not consistent, so it is defined as a contradiction point; according to the above table, all the contradiction points of each "voter" are found, which are counted as (N1, N2, …, N7).

[0110] A B C E F H D A P P N N P P N B P P P N N P P C P P P N N P N E N N P P N P P F N N N P P P N H N P P P P P P D P P N P N P P

[0111] Table 1: specific voting sequence in sports field

[0112] Finally, according to the "contradiction point", the credibility of the voter is measured, if the "contradiction point" is less than the preset value, it is considered that the user has high credibility, if the "contradiction point" is not less than the preset value, it is considered that the user has low credibility.

[0113] Specifically, according to the evaluation of the credibility of the voters, the correction of the voting is carried out, and the votes of the "voters" with low credibility are flipped over, so as to obtain the final voting sequence. If the final voting sequence is different from the voting sequence of the attitude of the four-tuple obtained by the voting method of the preset comment processing algorithm, it is determined that the comment information is the preset type of information, and if the final voting sequence is the same as the voting sequence of the attitude of the four-tuple obtained by the voting method of the preset comment processing algorithm, it is determined that the comment information is not the preset type of information.

[0114] In this embodiment, taking the identification of "anti-black" information in the sports field as an example, the embodiment is illustrated. The "anti-black" in other Baidu Bar and forum is likely to be "anti-black", which will cause the voters to cast "fake votes" and thus affect the final result, therefore, the distributed consistency algorithm PBFT will be modified and applied to the credibility of each voter, and the credibility will be determined by the comment attitude of the voter in other voters, so that the votes of the voters with credibility lower than the threshold can be reversed, that is, the originally cast N votes are changed to P votes, so as to avoid "fake votes" and obtain accurate results.

[0115] As Figure 8 The implementation process of the specific information identification method provided by the embodiment of the application is shown in the figure:

[0116] The first layer is to configure "anti-black" speech templates, when a user publishes a speech, if the speech matches the templates, the user is directly determined as an "anti-black" user, and the speech deletion and / or user speech ban operation is performed;

[0117] The second layer is to use the pre-trained "anti-black" speech discrimination model in the second layer to discriminate if the "anti-black" speech templates are not matched in the first layer, if the discrimination model considers it as "anti-black" speech, the speech deletion and / or speech ban operation is performed;

[0118] The third layer is to use the "anti-black" speech verification scoring algorithm to finally discriminate when the speech discrimination model cannot discriminate, if the speech is discriminated as "anti-black" speech, the speech deletion and / or speech ban operation is performed, otherwise, the user is considered as not an "anti-black" user.

[0119] The "anti-black" user identification method is illustrated by taking the "anti-black" user identification in the sports field as an example Figure 8 The specific "anti-black" user identification method is shown in the figure, for example:

[0120] Forum, Baidu post bar user published comments "certain day god, omnipotent, the world king, sweep the Champions League, " and the fact is that the team in the Champions League on the court defeat, this seemingly praise certain certain comments for the "anti string black" comments; when the user publishes this comment, first, the first layer of the comment template matching, if the comment template is configured in "the world king, " "king down, etc. Similar comments, directly match the "anti string black" comments, the user is defined as "anti string black" user; if not, then use the model to identify, if the model is identified as "anti string black" comments, the user is defined as "anti string black" user, the relevant operation; if the model also does not identify (i.e. the model does not identify as "anti string black" comments) then use the comment test scoring algorithm to identify, the algorithm is mainly considered in the relevant field of all comments of the user, in the example, certain is the sports field, the relevant field here can be "B post bar", "C post bar", "F post bar" and other sports post bar, on the basis of understanding the relevant field, that is, the user's comments on these other post bar semantic derivation of the user whether it is "anti string black" user, if so, then the relevant processing, if not, it is considered that the user's comments are not "anti string black" comments, the comment can be normally published.

[0121] The three-layer structure of "anti string black" user identification method can effectively identify the "anti string black" user who is good at disguising. First, a kind of comment template matching method combined with semantic is designed, which expands the matching range of comment template; second, a kind of comment attitude identification model based on improved GAN is designed, which can consider the user's comments and replies comprehensively to obtain the first attitude; then, the comment attitude identification model and the user comment identification model are used in combination, so that it can identify certain "anti string black" comments; then, the related field of user comment information is extracted according to the knowledge graph; finally, a kind of comment checking voting method is designed to comprehensively consider the user's comments to obtain the final identification result.

[0122] Figure 9 The specific "anti string black" user identification method is shown in the system Figure 8 The main module diagram of the system where the specific "anti string black" user identification method is shown is as follows Figure 9As shown, the system in which the "anti-blacklisting" user identification method is located is divided into an offline part and an online part. The offline part contains a training module of the speech discrimination model and a related field data acquisition module. The modules in this part are mainly used for offline regular operation, such as regular retraining of the model and crawling of data. The results generated are provided to the online module. The online part contains a user speech publishing module, a speech template configuration and matching module, a speech discrimination model identification module, a speech verification scoring module, an "anti-blacklisting" speech processing module, and a normal speech processing module. This part mainly implements the three-layer user speech discrimination method designed for the "anti-blacklisting" user identification system. The specific functions of each module are as follows:

[0123] User speech publishing module: This module mainly provides basic functions for users to publish speeches, such as functions for users to publish and edit their own speeches.

[0124] Speech template configuration and matching module: This module is first configured by an administrator with "anti-blacklisting" speech templates. When a user needs to publish a speech, the module performs semantic matching between the user's speech and the speech templates.

[0125] Speech discrimination model module: This module mainly discriminates whether the user's speech that does not match the speech templates is "anti-blacklisting" speech based on the trained speech discrimination model. This module contains two models. One is a speech attitude discrimination model used to discriminate whether the user's speech is positive praise or negative abuse. The other is a speech content discrimination model used to judge whether the speech is "anti-blacklisting" speech based on the discrimination result of the first model. The first model is also used in the speech verification scoring module.

[0126] Speech verification scoring module: This module is mainly responsible for judging whether the user's speech is "anti-blacklisting" speech based on the pre-set speech processing algorithm. The speech attitude discrimination model is used to discriminate various data of the user's speech extracted by the related field data acquisition module to support the speech verification scoring algorithm.

[0127] Discrimination model training module: The main function is to train and update the speech attitude discrimination model and the user speech discrimination model according to the set period, update the newly trained model to the speech discrimination model module, and provide the speech attitude discrimination model to the speech verification scoring module.

[0128] A related field data acquisition module, which is mainly used for acquiring the speech of a user in a forum or a post bar related to a field, and acquiring related field knowledge through a knowledge graph. For example, a user posts some speech in a post bar A, and according to the related association in the knowledge graph, it can be concluded that this is a sports post bar, and it can be obtained that post bar C, post bar B and post bar A are in an 'adversarial' relationship. Not only the data of these forums and post bars can be extracted, but also the extracted data can be pre-labeled for subsequent speech checking and scoring algorithms.

[0129] Anti-stringing black speech processing module, which deletes and / or mutes the speech of the identified user according to the identification result of the 'anti-stringing black' user.

[0130] Normal speech processing module, which does not process the normal speech of the user.

[0131] The above embodiments of the present application can accurately identify the speech information of the user through a three-layer identification model. The three-layer identification model also provides an identification method from three angles to make up for the respective shortcomings, and the accuracy of the identification is higher than that of a single use. The three methods all combine semantics, greatly reducing the participation of human beings, saving labor costs and making up for the shortcomings of manual review.

[0132] In addition, the speech attitude identification model considers the speech of the user and the reply speech of other users before identifying the first attitude of the user, so that the speech attitude identification model is more accurate. The speech checking and scoring model also considers the speech and attitude of the user in multiple related fields, and improves the PBIX algorithm to enable it to determine whether the user is an 'anti-stringing black', has better accuracy and makes up for the shortcomings of the existing model.

[0133] Figure 10 A structural schematic diagram of an information identification device 100 provided by an embodiment of the present application is shown. As shown in FIG. 10, the device includes:

[0134] An acquisition module 101 is configured to acquire speech information posted by a user.

[0135] A processing module 102 is configured to identify the speech information to obtain a first identification result of whether the speech information is information of a preset type. If the first identification result indicates that the speech information is not the information of the preset type, the user credibility is confirmed according to the speech information, and a second identification result of whether the speech information is the information of the preset type is confirmed according to the user credibility.

[0136] Optionally, the processing module 102, in identifying the speech information to obtain a first identification result of whether the speech information is information of a preset type, is specifically configured to:

[0137] matching the speech information with a preset speech template, if the matching is successful, determining that the speech information is information of a preset type; if the matching is unsuccessful, inputting the speech information into a trained preset speech discrimination model for processing to output the first identification result of whether the speech information is information of a preset type.

[0138] Optionally, matching the speech information with a preset speech template, if the matching is successful, determining that the speech information is information of a preset type, comprises:

[0139] converting the speech information into a semantic vector by using a preset semantic conversion model;

[0140] querying, according to the semantic vector, a semantic index of a semantic vector library stored in advance, if there is at least one speech template with a similarity to the speech information greater than a preset similarity threshold, then the matching is successful, and it is determined that the speech information is information of a preset type.

[0141] Optionally, inputting the speech information into a trained preset speech discrimination model for processing to output the first identification result of whether the speech information is information of a preset type, comprises:

[0142] inputting the speech information into a speech attitude discrimination model of the trained preset speech discrimination model for processing to obtain a first processing result; the first processing result comprises a triple <speech information published by a user, speech reply data, attitude>;

[0143] inputting the first processing result into a speech content discrimination model of the trained preset speech discrimination model for processing to output the first identification result of whether the speech information is information of a preset type.

[0144] Optionally, inputting the speech information into a speech attitude discrimination model of the trained preset speech discrimination model for processing to obtain a first processing result, comprises:

[0145] converting the speech information into a binary tuple <speech information published by a user, speech reply data>;

[0146] inputting the speech information published by the user into a first adversarial network G1 of the speech attitude discrimination model of the trained preset speech discrimination model for processing to obtain a first output, the first adversarial network G1 inputs the speech information published by the user into a first discriminator D1 to generate first target speech reply data;

[0147] The speech reply data is input into a second adversarial network G2 of the speech attitude discrimination model of the trained preset speech discrimination model for processing to obtain a second output, and the second adversarial network G2 inputs the speech reply data into a second discriminator D2 to generate first target speech information;

[0148] The first output and the second output are input into a long short-term memory (LSTM) layer of the speech attitude discrimination model of the trained preset speech discrimination model for processing to obtain an output of the LSTM layer;

[0149] The first target speech reply data and the first target speech information are input into a third discriminator D3 of the speech attitude discrimination model of the trained preset speech discrimination model, and the output of the LSTM layer is input into the third discriminator D3 to obtain the first processing result.

[0150] Optionally, the first processing result is input into the speech content discrimination model of the trained preset speech discrimination model for processing to output a first recognition result of whether the speech information is information of a preset type, including:

[0151] The speech information input by a user is input into a first encoder of the speech content discrimination model of the trained preset speech discrimination model for processing to obtain a first encoding vector;

[0152] The speech reply data is input into a second encoder of the speech content discrimination model for processing to obtain a second encoding vector;

[0153] The attitude is input into a third encoder of the speech content discrimination model for processing to obtain a third encoding vector;

[0154] The first encoding vector, the second encoding vector, and the third encoding vector are input into a convolution layer of the speech content discrimination model for processing to obtain a convolution output;

[0155] The convolution output is input into an enhancement layer of the speech content discrimination model for processing to obtain an enhancement output;

[0156] The enhancement output is input into a fully connected layer of the speech content discrimination model for processing to output a first recognition result of whether the speech information is information of a preset type.

[0157] Optionally, the processing module 102 is specifically configured to:

[0158] According to the opinion information, relevant knowledge information is acquired from a knowledge graph;

[0159] According to the relevant knowledge information, attitude information of the user is determined;

[0160] According to the attitude of the user, a voting algorithm is used to determine a relationship matrix of voters;

[0161] According to the relationship matrix of the voters and the attitude information of the user, user credibility is determined;

[0162] According to the user credibility, a second recognition result of whether the opinion information is information of a preset type is determined.

[0163] It should be noted that the embodiment is a device embodiment corresponding to the above-mentioned method embodiment, and all implementation manners in the above-mentioned method embodiment are applicable to the device embodiment, and the same technical effects can be achieved.

[0164] The embodiment of the application provides a nonvolatile computer storage medium, the computer storage medium stores at least one executable instruction, and the computer executable instruction can execute the information recognition method in any method embodiment.

[0165] Figure 11 The structure of the computing device provided by the embodiment of the application is shown, and the specific implementation of the computing device is not limited in the embodiment of the application.

[0166] As shown in Figure 11 The computing device can include a processor, a communications interface, a memory, and a communications bus.

[0167] The processor, the communications interface, and the memory complete mutual communication through the communications bus. The communications interface is used for communication with network elements such as clients or other servers. The processor is used for executing a program, and can specifically execute related steps in the information recognition method embodiment for the computing device.

[0168] Specifically, the program can include program code, and the program code includes computer operation instructions.

[0169] The processor can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement one or more embodiments of the application. The one or more processors included in the computing device can be of the same type or different types. For example, the one or more processors can include one or more CPUs and one or more ASICs.

[0170] The memory is configured to store a program. The memory can include a high speed RAM memory and can also include a non-volatile memory, such as at least one disk memory.

[0171] The program can be specifically configured to cause the processor to perform the information identifying method in any of the above method embodiments. The specific implementation of each step in the program can refer to the corresponding description in the corresponding step and unit of the information identifying method embodiments described above, and will not be described here. It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, the specific working process of the devices and modules described above can refer to the corresponding process description in the foregoing method embodiments, and will not be described here.

[0172] The algorithms and displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems can be used with these teachings, or with modifications that are within the scope of the application. The structure required to be present on such systems to implement the operations described above is apparent from the description above. In addition, the embodiments of the application are not described with reference to any particular programming language. It will be appreciated that a variety of programming languages can be used to implement the teachings of the embodiments of the application as described herein, and any references below to specific languages are provided for disclosure of the best mode of practicing the embodiments of the application.

[0173] In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the application can be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been described in detail in order not to obscure the understanding of this description.

[0174] Similarly, it is to be understood that the embodiments of the application can be alternately or additionally described employing the following terminology: blocks, components, elements, features or steps which are either individually recited or described in the disclosure or enumerated in any combination of two or more blocks or clauses in the claims and / or specification are intended to be interpreted independently in their respective contexts. It will be understood that any described features can be replaced by alternative features serving the same, equivalent, or a similar purpose, unless expressly stated otherwise. By way of example, any of the disclosed or recited features can be explicitly, implicitly or explicitly applied to some, each and every aspect of the specification as well as any combination of aspects, without restriction. While compositions and methods are described in terms of "comprising," "containing," or "including" various components or steps, the compositions and methods can also "consist essentially of" or "consist of" the various components and steps. All patents and publications mentioned herein are hereby incorporated by reference in their entirety.

[0175] Those skilled in the art will appreciate that the modules in the apparatuses in the embodiments can be self-adaptively changed and disposed in one or more apparatuses different from the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and furthermore can be divided into multiple sub-modules or sub-units or sub-components. Except that at least some of such features and / or processes or units are mutually exclusive, all combinations of all features disclosed in this specification (including the accompanying claims, abstract and drawings) and all processes or units of any methods or apparatuses disclosed so far can be adopted. Unless explicitly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract and drawings) can be replaced by an alternative feature serving the same, equivalent or similar purpose.

[0176] Furthermore, those skilled in the art will appreciate that a combination of features of different embodiments is intended to be within the scope of the application and forms a different embodiment. For example, in the following claims, any one of the claimed embodiments can be used in any combination.

[0177] The various component embodiments of the present application can be implemented in hardware, or as software modules running in one or more processors, or in combinations thereof. Skilled persons should appreciate that a microprocessor or a digital signal processor (DSP) can be used in practice to implement some or all of the functions of some or all of the components according to the embodiments of the present application. The embodiments of the present application can also be implemented as a program for executing part or all of the methods described herein on a computer or a processor (for example, a computer program and a computer program product). Such a program implementing the embodiments of the present application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or in any other form.

[0178] It should be noted that the above-mentioned embodiments illustrate rather than limit the application, and that those skilled in the art will be able to design many alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word 'comprising' does not exclude the presence of elements or steps other than those listed in a claim. The word 'a' or 'an' preceding an element does not exclude the presence of a plurality of such elements. The application can be implemented by means of both hardware and software, and any combinations of them. In a unit claim, any reference signs placed between parentheses in the claim do not limit the claim. The use of the word 'at least' followed by a list of one or more items does not exclude additional such items. The use of the words 'one' or 'the' to refer to an element of a claim does not exclude the presence of a plurality of such elements. The implementation of the method steps described in the claims to produce the items recited in the claims is not limited to the time or place at which the respective step is described, but can be performed in any order or simultaneously, and at least in part concurrently with each other.

Claims

1. An information recognition method characterized by comprising: The method comprises the following steps: obtaining comment information published by a user; identifying the comment information to obtain a first identification result of whether the comment information is information of a preset type; wherein the comment information is converted into a binary tuple <comment information published by a user, speech reply>, the binary tuple is input into a speech attitude identification model of a trained preset speech identification model to obtain a ternary tuple <comment information published by a user, speech reply, attitude>, and the ternary tuple <comment information published by a user, speech reply, attitude> is input into the trained preset speech identification model to obtain the first identification result; if the first identification result indicates that the comment information is not the information of the preset type, confirming a user credibility according to the comment information, and obtaining a second identification result of whether the comment information is the information of the preset type according to the user credibility; wherein the user credibility is determined according to a relationship matrix of voters and attitude information of the user; the attitude information of the user is determined according to relevant knowledge information obtained from a knowledge graph; and the relationship matrix of the voters is determined by using a voting algorithm according to the attitude information of the user.

2. The information recognition method according to claim 1, characterized by, identifying the comment information to obtain a first identification result of whether the comment information is information of a preset type, comprising: matching the comment information with a preset speech template, if the matching is successful, determining that the comment information is the information of the preset type, and if the matching is unsuccessful, inputting the comment information into a trained preset speech identification model to process and output the first identification result of whether the comment information is the information of the preset type.

3. The information recognition method according to claim 2, characterized by, matching the comment information with a preset speech template, if the matching is successful, determining that the comment information is the information of the preset type, comprising: converting the comment information into a semantic vector by using a preset semantic conversion model; querying from a semantic index of a pre-stored semantic vector library according to the semantic vector, if there is at least one speech template with a similarity greater than a preset similarity threshold to the comment information, the matching is successful, and it is determined that the comment information is the information of the preset type.

4. The information recognition method according to claim 2, characterized by, inputting the comment information into a speech attitude identification model of a trained preset speech identification model to process and output the first identification result of whether the comment information is the information of the preset type, comprising: processing the comment information by inputting the comment information into the speech attitude identification model of the trained preset speech identification model to obtain a first processing result; the first processing result comprises a ternary tuple <comment information published by a user, speech reply data, attitude>; inputting the first processing result into a speech content identification model of the trained preset speech identification model to process and output the first identification result of whether the comment information is the information of the preset type.

5. The information recognition method according to claim 4, characterized by, inputting the comment information into a speech attitude identification model of a trained preset speech identification model to process and obtain a first processing result, comprising: converting the comment information into a binary tuple <comment information published by a user, speech reply data>; The speech information published by the user is input into a first adversarial network G1 of a speech attitude discrimination model of a preset speech discrimination model to obtain a first output, and the first adversarial network G1 inputs the speech information published by the user into a first discriminator D1 to generate first target speech reply data; The speech reply data is input into a second adversarial network G2 of the speech attitude discrimination model of the preset speech discrimination model to obtain a second output, and the second adversarial network G2 inputs the speech reply data into a second discriminator D2 to generate first target speech information; The first output and the second output are input into a long short-term memory (LSTM) layer of the speech attitude discrimination model of the preset speech discrimination model to obtain an output of the LSTM layer; The first target speech reply data and the first target speech information are input into a third discriminator D3 of the speech attitude discrimination model of the preset speech discrimination model, and the output of the LSTM layer is input into the third discriminator D3 to obtain the first processing result.

6. The information recognition method according to claim 4, characterized by, The first processing result is input into a speech content discrimination model of the preset speech discrimination model to output a first identification result of whether the speech information is information of a preset type, including: The speech information published by the user of the triple <speech information published by the user, speech reply data, attitude> is input into a first encoder of a speech content discrimination model of a preset speech discrimination model to obtain a first encoding vector; The speech reply data is input into a second encoder of the speech content discrimination model to obtain a second encoding vector; The attitude is input into a third encoder of the speech content discrimination model to obtain a third encoding vector; The first encoding vector, the second encoding vector, and the third encoding vector are input into a convolution layer of the speech content discrimination model to obtain a convolution output; The convolution output is input into an enhancement layer of the speech content discrimination model to obtain an enhanced output; The enhanced output is input into a fully connected layer of the speech content discrimination model to output the first identification result of whether the speech information is information of a preset type.

7. The information recognition method according to Claim 1, characterized by, According to the speech information, the user credibility is confirmed, and according to the user credibility, a second identification result of whether the speech information is information of a preset type is confirmed, including: According to the speech information, related knowledge information is obtained from a knowledge graph; According to the related knowledge information, the attitude information of the user is determined; According to the attitude information of the user, a voting algorithm is used to determine a relationship matrix of voters; According to the relationship matrix of the voters and the attitude information of the user, the user credibility is determined; According to the user credibility, the second identification result of whether the speech information is information of a preset type is determined.

8. An information recognizing apparatus characterized by comprising: The device includes: An acquisition module configured to acquire speech information published by a user; The processing module is configured to identify the comment information to obtain a first identification result of whether the comment information is information of a preset type; wherein the comment information is converted into a binary tuple <comment information published by a user, speech reply>, the binary tuple is input into a comment attitude identification model of a trained preset comment identification model to obtain a ternary tuple <comment information published by a user, speech reply, attitude>, the ternary tuple <comment information published by a user, speech reply, attitude> is input into the trained preset comment identification model for processing to obtain the first identification result; if the first identification result indicates that the comment information is not the information of the preset type, a user credibility is confirmed according to the comment information, and a second identification result of whether the comment information is the information of the preset type is confirmed according to the user credibility; wherein the user credibility is determined according to a relationship matrix of voters and attitude information of the user; the attitude information of the user is determined according to relevant knowledge information obtained from a knowledge graph; and the relationship matrix of the voters is determined by using a voting algorithm according to the attitude information of the user.

9. A computing device comprising: A processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface complete communication with each other through the communication bus; The memory is configured to store at least one executable instruction, and the executable instruction causes the processor to perform operations corresponding to the information identification method in any one of claims 1-7.

10. A computer storage medium, the storage medium storing at least one executable instruction, the executable instruction causing a processor to perform operations corresponding to the information identification method in any one of claims 1-7.

Images