A real-time intelligent identification method and system for game chat advertisements

Abstract

The present invention discloses a real-time intelligent identification method and system for game chat advertisements. The chat feature data of each player is extracted from the game chat content of the game server in real time, and the chat feature data is distributed to a group of chat risk control servers for processing through the load balancing architecture of the ngnix server and the asynchronous mechanism of the distributed database redis; the chat feature data refreshed to the redis on the group of chat risk control servers is asynchronously processed through a plurality of advertisement identification rules / models set in each of the group of chat risk control servers, and whether there is advertisement content is identified in real time. Compared with the existing game chat advertisement identification scheme, the technical scheme provided by the present invention has good stability, fast processing speed, strong real-time performance, high recognition accuracy and high automation.

Description

Technical Field

[0001] The solution provided by the present invention relates to the field of text processing and semantic recognition, and is used to identify / detect content with specific semantic tendencies in text; specifically, it relates to a real-time intelligent recognition method and system for game chat advertisements. Background Art

[0002] There are often some "players" in the game who promote items or other services, as well as advertising players who attempt to divert users to other games. These players are often professional, and directly or "implicitly" promote items / equipment to other players during the game. If these promotional / advertising behaviors are allowed, it may cause losses to the game operator or other players, and also affect the player experience during the game.

[0003] At present, in order to identify advertising content in game chats, manual recognition or deep neural network algorithm-based recognition models are usually used to match chat texts. For example, the traditional data mining model inputs each player's real-time speech text into the random forest model to determine whether the player's speech is illegal. The features used by the model are all derived from manual extraction. Existing game chat advertising recognition solutions often have slow recognition speeds, poor real-time performance, and are not stable enough. In addition, due to the single recognition model, the conventional recognition model cannot accurately identify advertising content in chats, especially in multi-sentence texts that can only reflect the nature of "advertisements". The text recognition ability is insufficient. Summary of the invention

[0004] In order to solve the above-mentioned deficiencies of the existing game chat advertisement recognition scheme, the present invention provides a real-time intelligent recognition scheme for game chat advertisements. The scheme is based on a distributed storage system and a real-time computing engine to obtain the chat data of each player in the game chat in real time, and refreshes the extracted chat data to a group of chat risk control servers in distributed redis in real time through load balancing technology for multi-rule / model advertisement recognition. The scheme has good stability, fast processing speed, strong real-time performance, and high recognition accuracy and automation.

[0005] The technical solution provided by the present invention is specifically implemented as follows:

[0006] A real-time intelligent identification method for game chat advertisements, the method comprising: extracting chat feature data of players from the game chat content of a game server in real time, distributing the chat feature data to a group of chat risk control servers for processing through the load balancing architecture of the ngnix server and the asynchronous mechanism of the distributed database redis; asynchronously processing the chat feature data refreshed to the redis on the chat risk control server through multiple advertisement identification rules / models set on the chat risk control server, and identifying in real time whether there is advertisement content. The chat feature data of different players extracted in real time from the game chat content at least includes: role, IP, and speech text during the chat.

[0007] There are two paths for players’ chat feature data to travel from the game server to the chat risk control server:

[0008] 1. For chat features that do not need to be calculated in combination with the player's historical information, the stream processing platform kafka is used to send them directly to the ngnix server, and the ngnix server is used to perform a load balancing mechanism to send the chat feature data to the redis in the group of chat risk control servers.

[0009] 2. For chat feature data related to player historical data that can only be obtained by combining player historical data, the stream processing platform kafka sends the relevant chat data to the distributed storage cassandra or kudu used to store player historical data for storage; the real-time computing engine presto calculates the chat feature data related to the player historical data corresponding to the specific player based on all relevant data of the specific player in the distributed storage cassandra or kudu, and then writes it into the search analysis engine es (Elastic Search), which is refreshed by es into the ngnix server in real time, and the chat feature data related to the player historical data is sent to the redis in the group of chat risk control servers through the load balancing mechanism of the ngnix server.

[0010] Furthermore, the multiple advertising recognition rules / models include: whitelist matching rules based on player information, blacklist matching rules based on player information, sensitive words and rules formulated by long-term experience of game operators, single-text semantic advertising recognition model based on multi-head self-attention mechanism, multi-text semantic advertising recognition model based on Word2Vec+BiLSTM architecture and random forest advertising recognition model with real-time feature input.

[0011] Among them, the single text semantic advertisement recognition model is a multi-head self-attention architecture based on the transform architecture, and is trained using a large number of manually annotated positive and negative samples. The advertisement recognition model includes: a preprocessing layer for cleaning the speech text to be recognized and inputting it into the word embedding layer; a digital feature matrix for extracting the input text and a position matrix for recording the position of each word in the text. The word2vec model used in this layer is obtained by learning a large number of in-game historical chat sessions collected by operation and maintenance personnel using the CBOW or Skip-Gram method; a multi-layer perceptron MPL composed of a cascade of processing units composed of a multi-head self-attention module and an FFN layer, which is used to receive the digital matrix and the position matrix for processing, and obtain the text feature matrix related to the text attention and input it into the fully connected layer; and a fully connected layer for mapping the text feature matrix to the sample tag space and predicting whether the text is an advertisement content. The fully connected layer is a multi-layer structure, and its last layer outputs a value that is processed by the sigmoid function to generate the probability of whether the text is an advertisement.

[0012] The multi-text semantic advertisement recognition model is based on the Word2Vec+BiLSTM deep learning architecture and is trained using a large number of manually annotated positive and negative samples. It includes a preprocessing layer, a word embedding layer, a normalization layer layer norm, a bidirectional long short-term memory layer, and a fully connected layer. Among them, the preprocessing layer is used to clean the multi-sentence text formed by splicing together the recent predetermined number of speech texts of a specific player and input it into the word embedding layer; the word embedding layer is used to extract the digital feature matrix of the input text and the position matrix that records the position of each word in the text, and input the digital feature matrix and the position matrix into the normalization layer layer norm for standard normalization processing. The word2vec model used in the word embedding layer is obtained by learning a large number of in-game historical chat sessions collected by operation and maintenance personnel using the CBOW or Skip-Gram method; the bidirectional long short-term memory layer BiLSTM is used to extract the context information matrix of the context dependency information of the multi-sentence text and output it to the fully connected layer to predict whether the multi-sentence text is an advertisement.

[0013] Furthermore, the chat risk control server identifies chat ads in a progressive manner based on the multiple ad recognition rules / models. The specific process is: first, a whitelist matching rule based on player information is used to match a player's chat data for the first time. If the player information matches the information in the whitelist, it is directly passed without subsequent detection. Otherwise, the chat data is matched for the second time using the blacklist matching rule. When the player is in the blacklist, or his chat speech text contains sensitive words in the blacklist, or meets the rules formulated by the long-term experience of the game operators in the blacklist, an alarm is directly triggered. Otherwise, the single text semantic ad recognition model is used to identify ads in the speech text in the chat data; if the recognition result of the single text semantic ad recognition model based on the multi-head attention mechanism is that it belongs to an ad, an alarm is triggered directly. Otherwise, the player's recent predetermined number of speech texts are spliced ​​together to form a multi-sentence text, and the multi-text semantic ad recognition model is used for ad recognition. Preferably, for multi-sentence texts predicted as non-advertisements by the multi-text semantic advertising recognition model, a traditional data mining model is further used to input the chat data of each player into a random forest model, and output whether the player is a violator to improve the recall rate of the system; the features used in the random forest model are all derived from manual extraction, and the decision-making process is scientific and reliable.

[0014] Corresponding to the above method, the present invention also provides a real-time intelligent recognition system for game chat advertisements. The system includes: a stream processing platform kafka, a distributed storage system cassandra or kudu, a real-time computing engine presto, a group of chat risk control servers, ngnix servers and a search analysis engine es;

[0015] Among them, for chat feature data that can be obtained without combining with the player's historical information, the stream processing platform kafka directly sends it to the ngnix server, and uses the ngnix server to perform a load balancing mechanism to send the chat feature data to the redis in the group of chat risk control servers; for chat feature data related to the player's historical data that can only be obtained by combining with the player's historical data, the stream processing platform kafka sends the relevant chat data to the distributed storage cassandra or kudu for storing the player's historical data, and the real-time computing engine presto calculates the chat feature data related to the player's historical data corresponding to the specific player according to all the relevant data of the specific player in the distributed storage cassandra or kudu, and then writes it into the search analysis engine es (ElasticSearch), which is refreshed by es into the ngnix server in real time, and the chat feature data related to the player's historical data is sent to the redis in the group of chat risk control servers through the load balancing mechanism of the ngnix server;

[0016] The real-time computing engine Presto extracts chat data of different players from the game chat content in real time and stores it in the distributed storage system Cassandra;

[0017] The ngnix server communicates with the distributed storage system cassandra, and based on its own load balancing architecture and through the asynchronous mechanism of distributed redis, distributes the chat data to the group of chat risk control servers for processing;

[0018] Each of the group of chat risk control servers is provided with a plurality of advertisement identification rules / models, and asynchronously processes the chat data refreshed to the redis thereon based on the plurality of advertisement identification rules / models, and identifies in real time whether there is advertisement content.

[0019] The specific implementation process of the above-mentioned real-time intelligent recognition system for game chat advertisements corresponds to the details of the above-mentioned real-time intelligent recognition method for game chat advertisements. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of multiple advertisement identification rules / models set in the chat risk control server in the solution provided by the present invention. DETAILED DESCRIPTION

[0021] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0022] A real-time intelligent identification method for game chat advertisements, the method comprising: extracting chat feature data of each player from the game chat content of a game server in real time, distributing the chat feature data to a group of chat risk control servers for processing through the load balancing architecture of an ngnix server and the asynchronous mechanism of a distributed database redis; asynchronously processing the chat feature data refreshed to redis on the chat risk control server through a variety of advertisement identification rules / models set on the chat risk control server, and identifying in real time whether there is advertisement content. The chat feature data of each player extracted in real time from the game chat content at least includes: role, IP, and speech text during the chat.

[0023] There are two paths for players’ chat feature data to travel from the game server to the chat risk control server:

[0024] 1. For chat features that do not need to be calculated in combination with the player's historical information, the stream processing platform kafka is used to send them directly to the ngnix server, and the ngnix server is used to perform a load balancing mechanism to send the chat feature data to the redis in the group of chat risk control servers.

[0025] 2. For chat feature data related to player historical data that can only be obtained by combining player historical data, the stream processing platform kafka sends the relevant chat data to the distributed storage cassandra or kudu used to store player historical data for storage; the real-time computing engine presto calculates the chat feature data related to the player historical data corresponding to the specific player based on all relevant data of the specific player in the distributed storage cassandra or kudu, and then writes it into the search analysis engine es (Elastic Search), which is refreshed by es into the ngnix server in real time, and the chat feature data related to the player historical data is sent to the redis in the group of chat risk control servers through the load balancing mechanism of the ngnix server.

[0026] Among them, Cassandra is an open source distributed NoSQL database system. It was originally developed by Facebook to store simple format data such as inboxes. It combines the data model of Google BigTable with the fully distributed architecture of Amazon Dynamo. Facebook open-sourced Cassandra in 2008. Since then, due to its good scalability, Cassandra has been adopted by well-known Web 2.0 websites such as Digg and Twitter, becoming a popular distributed structured data storage solution. Cassandra is a hybrid non-relational database, similar to Google's BigTable. Its main functions are richer than Dynamo (distributed Key-Value storage system), and it is the most functional and relational database among non-relational databases. The supported data structure is very loose, in the bjson format similar to json, so it can store more complex data types. The Presto query engine is a Master-Slave architecture, consisting of a Coordinator node, a Discovery Server node, and multiple Worker nodes. The Discovery Server is usually embedded in the Coordinator node. The Coordinator is responsible for parsing SQL statements, generating execution plans, and distributing execution tasks to Worker nodes for execution. Due to the above characteristics, it is very suitable for cooperation with the distributed storage system Cassandra to realize the capture, storage and query of each player's chat data.

[0027] Remote dictionary service (Redis) is a key-value storage system that stores data in memory, supports incremental data operations, avoids full data operations when querying data, and supports rich data operations such as push / pop, add / remove, etc., and these operations are atomic, avoiding data inconsistency problems. Ngnix server itself has a built-in load balancing module, which can facilitate load balancing of data distribution according to different application scenarios.

[0028] like Figure 1 As shown in the technical solution provided by the present invention, the multiple advertisement identification rules / models in the chat risk control server include:

[0029] The direct pass model mainly includes whitelist matching rules based on player information. For players who meet the rules and are on the IP and account whitelists, the system returns to normal directly.

[0030] Direct prohibition model, which contains blacklist matching rules based on player information, sensitive words and the rules formulated by the long-term experience of game operators. According to the rules formulated by the long-term experience of game operators, IP and account blacklists, when a player is on the blacklist or contains sensitive words in his speech or meets the rules, the system directly returns an alarm message.

[0031] A single text semantic advertisement recognition model based on a multi-head self-attention mechanism, the single text semantic advertisement recognition model is a multi-head self-attention architecture based on a transform architecture, and is trained using a large number of manually annotated positive and negative samples. The advertisement recognition model includes: a preprocessing layer for cleaning the speech text to be recognized and inputting it into the word embedding layer; a digital feature matrix for extracting the input text and a position matrix for recording the position of each word in the text, the word2vec model used in this layer is obtained by learning a large number of in-game historical chat sessions collected by operation and maintenance personnel using the CBOW or Skip-Gram method; a multi-layer perceptron MPL composed of a cascade of processing units composed of a multi-head self-attention module and an FFN layer, which is used to receive the digital matrix and the position matrix for processing, and obtain the text feature matrix related to the text attention and input it into the fully connected layer; and a fully connected layer for mapping the text feature matrix to the sample tag space and predicting whether the text is an advertisement content, the fully connected layer is a multi-layer structure, and the last layer outputs a value that is processed by a sigmoid function to generate the probability of whether the text is an advertisement. The single-text deep learning model based on the transform architecture can identify players who are not caught by the rules and blacklists through the semantics of their speeches, but whose speeches can be seen to have violated the rules based on the semantics.

[0032] A multi-text semantic advertisement recognition model is obtained by training based on the Word2Vec+BiLSTM deep learning architecture and a large number of manually annotated positive and negative samples. It includes a preprocessing layer, a word embedding layer, a normalization layer layer norm, a bidirectional long short-term memory layer, and a fully connected layer. The preprocessing layer is used to clean the multi-sentence text formed by splicing together a predetermined number of recent speeches of a specific player and input it into the word embedding layer; the word embedding layer is used to extract a digital feature matrix from the input text and a position matrix that records the position of each word in the text, and input the digital feature matrix and the position matrix into the normalization layer layer norm for standard normalization processing. The word2vec model used in the word embedding layer is obtained by learning a large number of in-game historical chat sessions collected by operation and maintenance personnel using the CBOW or Skip-Gram method; the bidirectional long short-term memory layer BiLSTM is used to extract the context information matrix of the context dependency information of the multi-sentence text and output it to the fully connected layer to predict whether the multi-sentence text is an advertisement. The advertisement recognition model is used to identify a player's speech within a period of time and determine whether the player is a player who makes illegal speeches, thereby greatly improving the recall rate of the system.

[0033] Traditional data mining model, the traditional data mining model is a random forest module, and the features used are all derived from manual extraction; the real-time chat data of each player is input into the random forest model to determine whether the player's speech is illegal.

[0034] The above-mentioned Word2Vec+BiLSTM deep learning architecture and the multi-head self-attention architecture based on the transform architecture are relatively mature architectures. The training process can adopt conventional training of technicians in this field, which will not be repeated here.

[0035] Whitelist matching rules based on player information, blacklist matching rules based on player information, sensitive words and the long-term experience of game operators, single-text semantic advertising recognition model based on multi-head self-attention mechanism, multi-text semantic advertising recognition model based on Word2Vec+BiLSTM architecture, and random forest advertising recognition model with real-time feature input.

[0036] Furthermore, in order to improve the accuracy of advertisement recognition, the chat risk control server identifies chat advertisements in a layered and progressive manner based on the multiple advertisement recognition rules / models. The specific process is as follows: first, a whitelist matching rule based on player information is used to perform a first match on a player's chat data. If the player information matches the information in the whitelist, it is directly passed without subsequent detection. Otherwise, the chat data is matched for a second time using the blacklist matching rule. When the player is in the blacklist, or his chat speech text contains sensitive words in the blacklist, or meets the rules formulated by the long-term experience of game operators in the blacklist, an alarm is directly triggered. Otherwise, the speech text in the chat data is used for advertisement recognition using the single-text semantic advertising recognition model; if the recognition result of the single-text semantic advertising recognition model based on the multi-head attention mechanism is that it belongs to an advertisement, an alarm is directly triggered. Otherwise, the player's most recent predetermined number of speech texts are spliced ​​together to form a multi-sentence text, and the multi-text semantic advertising recognition model is used for advertisement recognition; for the multi-sentence text predicted by the multi-text semantic advertising recognition model as non-advertisement, the chat data of each player is further input into the random forest model using a traditional data mining model, and output whether the player is in violation to improve the recall rate of the system.

[0037] Corresponding to the above method, the present invention also provides a real-time intelligent identification system for game chat advertisements, characterized in that the system includes: a stream processing platform kafka, a distributed storage system cassandra or kudu, a real-time computing engine presto, a group of chat risk control servers, an ngnix server and a search analysis engine es.

[0038] Among them, for chat feature data that can be obtained without combining with the player's historical information, the stream processing platform kafka directly sends it to the ngnix server, and uses the ngnix server to perform a load balancing mechanism to send the chat feature data to the redis in the group of chat risk control servers; for chat feature data related to the player's historical data that can only be obtained by combining with the player's historical data, the stream processing platform kafka sends the relevant chat data to the distributed storage cassandra or kudu for storing the player's historical data, and the real-time computing engine presto calculates the chat feature data related to the player's historical data corresponding to the specific player according to all the relevant data of the specific player in the distributed storage cassandra or kudu, and then writes it into the search analysis engine es (ElasticSearch), which is refreshed by es into the ngnix server in real time, and the chat feature data related to the player's historical data is sent to the redis in the group of chat risk control servers through the load balancing mechanism of the ngnix server;

[0039] The real-time computing engine Presto extracts chat data of different players from the game chat content in real time and stores it in the distributed storage system Cassandra;

[0040] The ngnix server communicates with the distributed storage system cassandra, and based on its own load balancing architecture and through the asynchronous mechanism of distributed redis, distributes the chat data to the group of chat risk control servers for processing;

[0041] Each of the group of chat risk control servers is provided with a plurality of advertisement identification rules / models, and asynchronously processes the chat data refreshed to the redis thereon based on the plurality of advertisement identification rules / models, and identifies in real time whether there is advertisement content.

[0042] The specific details of the real-time intelligent recognition system for game chat advertisements performing advertisement recognition on player chat data correspond to the details of the real-time intelligent recognition method for game chat advertisements performing advertisement recognition on player chat data.

[0043] The real-time intelligent identification scheme for game chat advertisements provided by the present invention cooperates with each other by combining a distributed storage system and a real-time computing engine to obtain the chat data of each player in the game chat in real time, thereby improving the speed of data extraction and meeting the requirement of real-time extraction of chat data in the game; the extracted chat data is refreshed in real time to a distributed redis on a group of chat risk control servers through load balancing technology for processing, thereby improving the stability of the system and preventing the chat risk control server from performing full operations on the chat data of the players. An advertisement recognition model including multiple rules / models including a deep learning model is set in the chat risk control server, and advertisement recognition of the player's chat data is performed by layer-by-layer detection, which greatly improves the recognition accuracy and high degree of automation.

Claims

1. A real-time intelligent identification method for game chat advertisements. It is characterized in that The method comprises: extracting chat feature data of each player from the game chat content of the game server in real time, distributing the chat feature data to a group of chat risk control servers for processing through the load balancing architecture of the ngnix server and the asynchronous mechanism of the distributed database redis; asynchronously processing the chat feature data refreshed to the redis on the group of chat risk control servers through a plurality of advertisement identification rules or models set in each of the group of chat risk control servers, and identifying in real time whether there is advertisement content; The multiple advertisement recognition rules or models are used to identify chat advertisements in a progressive manner, specifically including: firstly, a whitelist matching rule based on player information is used to perform a first match on a player's chat feature data. If the player information matches the information in the whitelist, it is directly passed without subsequent detection. Otherwise, a blacklist matching rule is used to perform a second match on the chat feature data. When the player is in the blacklist, or his chat speech text contains sensitive words in the blacklist, or meets the rules formulated by the long-term experience of the game operators in the blacklist, an alarm is directly triggered. Otherwise, a single text semantic advertisement recognition model based on a multi-head self-attention mechanism is used to perform advertisement recognition on the speech text in the chat feature data; if the recognition result of the single text semantic advertisement recognition model based on the multi-head self-attention mechanism is that it belongs to an advertisement, an alarm is directly triggered. Otherwise, a predetermined number of speech texts of the player are spliced ​​together to form a multi-sentence text, and the multi-text semantic advertisement recognition model is used to perform advertisement recognition; for multi-sentence texts identified as non-advertisements, a random forest advertisement recognition model with real-time feature input is used for recognition; The chat feature data from the game server to the chat risk control server includes two paths: a. For chat feature data that can be obtained without combining historical information of players, the stream processing platform kafka is used to send it directly to the ngnix server, and the load balancing mechanism of the ngnix server is used to send the chat feature data to the redis in the group of chat risk control servers; b. For chat feature data that needs to be calculated in combination with player historical data, the stream processing platform kafka is used to send the relevant chat content to the distributed storage cassandra or kudu used to store player historical data for storage; the real-time computing engine presto calculates the chat feature data corresponding to the specific player based on all relevant data of the specific player in the distributed storage cassandra or kudu, and then writes it to the search analysis engine es, and then the search analysis engine es refreshes this part of the chat feature data into the ngnix server in real time, and sends it to the redis of the group of chat risk control servers through the ngnix server.

2. The method according to claim 1, It is characterized in that The chat feature data of each player extracted in real time includes at least: role, IP, and speech text during the chat.

3. The method according to claim 2, It is characterized in that The multiple advertising recognition rules or models include: whitelist matching rules based on player information, blacklist matching rules based on player information, sensitive words and long-term experience of game operators, single-text semantic advertising recognition model based on multi-head self-attention mechanism, multi-text semantic advertising recognition model based on Word2Vec+BiLSTM architecture and random forest advertising recognition model with real-time feature input.

4. The method according to claim 3, It is characterized in that The single-text semantic advertisement recognition model based on the multi-head self-attention mechanism is a multi-head self-attention architecture based on the transform architecture, trained with a large number of manually annotated positive and negative samples; including: a preprocessing layer, a word embedding layer, a multi-layer perceptron MPL and a fully connected layer; The preprocessing layer is used to clean the speech text to be recognized and then input it into the word embedding layer; the word embedding layer is used to extract the digital feature matrix of the input text and the position matrix of each word in the text. The word2vec model used in this layer is obtained by learning a large number of historical chat sessions in the game collected by operation and maintenance personnel using the CBOW or Skip-Gram method; The multi-layer perceptron MPL is composed of a cascade of processing units consisting of a multi-head self-attention module and an FFN layer, and is used to receive the digital feature matrix and the position matrix for processing, and obtain the text feature matrix related to the text attention and input it into the fully connected layer to generate a prediction of whether the text is an advertisement; The fully connected layer is a multi-layer fully connected layer, which is used to map the text feature matrix to the sample tag space, and the last layer outputs a value which is processed by the sigmoid function to generate the probability of whether the text is an advertisement.

5. The method according to claim 3 or 4, It is characterized in that The multi-text semantic advertisement recognition model is based on the Word2Vec+BiLSTM deep learning architecture and is trained using a large number of manually annotated positive and negative samples; it includes: a preprocessing layer, a word embedding layer, a normalization layer layer norm, a bidirectional long short-term memory layer, and a fully connected layer; wherein the preprocessing layer is used to clean a multi-sentence text formed by splicing a predetermined number of speech texts of a specific player recently together and then input it into the word embedding layer; the word embedding layer is used to extract a digital feature matrix from the input text and a position matrix that records the position of each word in the text, and input the digital feature matrix and the position matrix into the normalization layer layernorm for standard normalization processing, and the word2vec model used in the word embedding layer is obtained by learning a large number of historical chat sessions in the game collected by operation and maintenance personnel using the CBOW or Skip-Gram method; the bidirectional long short-term memory layer BiLSTM is used to extract a context information matrix of context dependency information of the multi-sentence text and then output it to the fully connected layer to generate a prediction of whether the multi-sentence text is an advertisement.

6. A real-time intelligent recognition system for game chat ads, It is characterized in that The system includes: stream processing platform kafka, distributed storage system cassandra or kudu, real-time computing engine presto, a group of chat risk control servers, ngnix server and search analysis engine es; Among them, for chat feature data that can be obtained without combining with the player's historical information, the stream processing platform kafka directly sends it to the ngnix server, and uses the ngnix server to perform a load balancing mechanism to send the chat feature data to the redis in the group of chat risk control servers; for chat feature data related to the player's historical data that can only be obtained by combining with the player's historical data, the stream processing platform kafka sends the relevant chat content to the distributed storage cassandra or kudu used to store the player's historical data for storage, and the real-time computing engine presto calculates the chat feature data related to the player's historical data corresponding to the specific player based on all relevant data of the specific player in the distributed storage cassandra or kudu, and then writes it into the search analysis engine es, and the search analysis engine es refreshes the chat feature data related to the player's historical data into the ngnix server in real time, and sends it to the redis in the group of chat risk control servers through the load balancing mechanism of the ngnix server; The ngnix server communicates with the distributed storage system cassandra, and based on its own load balancing architecture and through the asynchronous mechanism of distributed redis, distributes the chat feature data to the group of chat risk control servers for processing; Each of the group of chat risk control servers is provided with a plurality of advertisement identification rules or models, and asynchronously processes the chat feature data refreshed to the redis thereon based on the plurality of advertisement identification rules or models, and identifies in real time whether there is advertisement content; The group of chat risk control servers uses a layered progressive approach to perform advertisement identification on the chat feature data refreshed thereon based on the multiple advertisement identification rules or models, specifically including: firstly, a whitelist matching rule based on player information is used to perform a first match on the chat feature data of a certain player; if the player information matches the information in the whitelist, it is directly passed without subsequent detection; otherwise, a blacklist matching rule is used to perform a second match on the chat feature data; when the player is in the blacklist, or his chat speech text contains sensitive words in the blacklist, or meets the rules formulated by the long-term experience of the game operators in the blacklist, an alarm is directly triggered; otherwise, a single text semantic advertisement recognition model based on a multi-head self-attention mechanism is used to perform advertisement identification on the speech text in the chat feature data; if the recognition result of the single text semantic advertisement recognition model based on the multi-head self-attention mechanism is that it belongs to an advertisement, an alarm is directly triggered; otherwise, a predetermined number of speech texts of the player are spliced ​​together to form a multi-sentence text, and the multi-text semantic advertisement recognition model is used to perform advertisement identification; for the multi-sentence text identified as non-advertisement, a random forest advertisement recognition model with real-time feature input is used for identification.

7. The system according to claim 6, It is characterized in that The multiple advertising recognition rules or models include: whitelist matching rules based on player information, blacklist matching rules based on player information, sensitive words and long-term experience of game operators, single-text semantic advertising recognition model based on multi-head self-attention mechanism, multi-text semantic advertising recognition model based on Word2Vec+BiLSTM architecture and random forest advertising recognition model with real-time feature input.

8. The system of claim 7, It is characterized in that The single text semantic advertisement recognition model based on the multi-head self-attention mechanism is a multi-head self-attention architecture based on the transform architecture, which is trained by using a large number of manually annotated positive and negative samples; it includes: a preprocessing layer, a word embedding layer, a multi-layer perceptron MPL and a fully connected layer; wherein, the preprocessing layer is used to clean the speech text to be recognized and then input it into the word embedding layer; the word embedding layer is used to extract the digital feature matrix of the input text and the position matrix of each word in the text. The word2vec model used in this layer is obtained by adopting CBOW or Skip. The p-Gram method is obtained by learning a large number of historical in-game chat sessions collected by operation and maintenance personnel; the multi-layer perceptron MPL is composed of a cascade of processing units consisting of a multi-head self-attention module and an FFN layer, which is used to receive the digital feature matrix and the position matrix for processing, and obtain the text feature matrix related to the text attention and input it into the fully connected layer to generate a prediction of whether the text is an advertisement. The fully connected layer is a multi-layer fully connected layer, which is used to map the text feature matrix to a sample label space, and its last layer outputs a value that is processed by a sigmoid function to generate a probability of whether the text is an advertisement.

9. The system according to claim 7 or 8, It is characterized in that The multi-text semantic advertisement recognition model is based on the Word2Vec+BiLSTM deep learning architecture and is trained using a large number of manually annotated positive and negative samples; it includes: a preprocessing layer, a word embedding layer, a normalization layer layer norm, a bidirectional long short-term memory layer, and a fully connected layer; The preprocessing layer is used to clean a multi-sentence text formed by splicing together a predetermined number of recent speech texts of a specific player and then input it into the word embedding layer; The word embedding layer, the digital feature matrix for extracting the input text, and the position matrix for recording the position of each word in the text, the digital feature matrix and the position matrix are input into the normalization layer layer norm for standard normalization processing, and the word2vec model used in the word embedding layer is obtained by learning a large number of historical chat sessions in the game collected by operation and maintenance personnel using the CBOW or Skip-Gram method; The bidirectional long short-term memory layer BiLSTM is used to extract the context information matrix of the context dependency information of the multi-sentence text and output it to the fully connected layer to generate a prediction of whether the multi-sentence text is an advertisement.

Images