A message intelligent recognition method and system for advertising AI customer service

Abstract

The present invention discloses a message intelligent recognition method and system for AI customer service of advertising. By converting words in sentences into characters, the features of words can be increased, and a first sentence detection network capable of controlling connection relationships by repeated words can be added. The features of a word can be extracted separately for repeated words at different positions in interactive sentences. When the network structure is the same, the corresponding features of repeated words should be the same, so the features can be fused through the repeated structure, and the fused features can be input into the arranged neurons according to the original interactive sentence sequence through the classification structure to extract the relationship between words. The detection of user demand keywords is made more accurate. And according to the repeated words in the sentence, according to the composition portrait of the user's interactive sentence, and the association relationship between the keywords and the advertising content, the corresponding advertisement corresponding to the advertising content that best suits the user is found together, thereby achieving the technical effect of more accurate advertising according to the needs of the user.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and system for intelligently identifying messages of AI customer service for advertising. Background Art

[0002] At present, with the continuous development of artificial intelligence technologies such as big data analysis, visual recognition, and speech recognition, people have put forward higher and higher requirements for intelligent interaction. When users interact with artificial intelligence, operations are provided by understanding user needs. At the same time, advertisements that are more in line with user needs can also be delivered through user interaction information. Generally, keywords mentioned by users are extracted for advertising, without judging based on the language state during user interaction, such as repeated words. And when the network is built, since the network has to meet the conditions of many different sentence inputs, the parameters of each position in the network are different. As a result, even when only features are extracted for individual words, although they are repeated words, the extracted features are different, so the detection is not accurate enough, the message recognition is inaccurate, and the system cannot accurately push advertising information. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for intelligent message recognition of AI customer service for advertising, so as to solve the above-mentioned problems existing in the prior art.

[0004] In a first aspect, an embodiment of the present invention provides a method for intelligently identifying messages of an AI customer service for advertising, including:

[0005] Obtaining an interactive sentence and a fixed interactive sentence length and a repeated word corresponding to the interactive sentence; the repeated word represents a repeated word in the interactive sentence;

[0006] Marking the repeated words according to their positions in the interactive sentences to obtain a repeated marking vector; the repeated marking vector includes n elements; the repeated marking vector marks the repeated words with the same number;

[0007] Obtaining a sentence detection network; the sentence detection network is a network capable of extracting keywords related to the pre-delivered advertisement in the interactive sentence;

[0008] Based on the repeated tag vector, adjusting the sentence convolutional network to obtain a first sentence detection network;

[0009] According to the interactive sentence, sentence recognition is performed through a first sentence detection network to obtain a demand keyword; the demand keyword indicates a target location expected by the user;

[0010] According to the fixed interactive sentence length, repeated words and required keywords, the associated advertising content is detected to obtain a first advertising label; the first advertising label represents the advertising information represented by the interactive sentence in the chat record.

[0011] Optionally, the first sentence detection network adds n storage structures, 1 repetition structure and 1 classification structure to the sentence detection network;

[0012] A neuron in the first hidden layer of the first sentence neural network is fully connected to m neurons in the input layer corresponding to a word in the interactive sentence;

[0013] The neurons of the first hidden layer of the first sentence neural network are connected to a storage structure; the storage structure indicates whether it is a feature corresponding to a repeated string;

[0014] n storage structures are fully connected to 1 repeating structure;

[0015] The repetition structure is connected to a classification structure; the classification structure stores the positions of repeated words;

[0016] The classification structure and the storage structure are fully connected with neurons of the second hidden layer;

[0017] The second hidden layer is fully connected to two adjacent hidden layers in a plurality of other hidden layers; the other hidden layers are hidden layers other than the first hidden layer and the second hidden layer;

[0018] The last hidden layer and the output layer are fully connected; the number of neurons in the output layer is equal to the fixed interaction sentence length.

[0019] Optionally, adjusting the sentence convolution network based on the repeated tag vector to obtain a first sentence detection network includes:

[0020] The storage structure includes storage features and storage values; the storage value is initially 0; the storage feature is a feature of a neuron output corresponding to a word in the interactive sentence in the first hidden layer;

[0021] According to the repeat mark vector, the storage value corresponding to the repeat word is set to 1; one element in the repeat mark vector corresponds to one storage structure and one storage value;

[0022] If the storage structure is 1, the storage feature is input into the repeating structure;

[0023] In the repetitive structure, the average value of multiple stored features corresponding to the same value in the repetitive tag vector is calculated to obtain a repetitive fusion feature; multiple labels in the repetitive tag vector correspond to multiple repetitive fusion features;

[0024] Based on the repeated tag vector and the repeated fusion feature, the classification structure is adjusted to obtain a first sentence detection network.

[0025] Optionally, adjusting the classification structure based on the repeated tag vector and the repeated fusion feature to obtain a first sentence detection network includes:

[0026] According to the order of each word in the interactive sentence, a classification vector is constructed in the classification structure according to the repeated mark vector; the positions corresponding to the same repeated words in the classification vector are set to 1, and the other positions are set to 0; multiple labels in the repeated mark vector correspond to multiple classification vectors;

[0027] According to the corresponding position in the classification vector, the repeated fusion feature is input into the corresponding neuron of the second hidden layer;

[0028] If the storage structure is 0, the storage feature is input into the corresponding neuron of the second hidden layer.

[0029] Optionally, performing sentence recognition through a first sentence detection network according to the interactive sentence to obtain the required keywords includes:

[0030] Match each word in the interactive sentence with a corresponding string to obtain a string set; one string corresponds to one string in the interactive sentence;

[0031] Obtaining a plurality of first string features corresponding to a plurality of strings in the string set;

[0032] Inputting a plurality of first character string features into a first sentence detection network to obtain a keyword vector; the length of the keyword vector is equal to the number of words in the sentence; an element in the keyword vector indicates whether the word in the sentence corresponding to the element is a keyword related to advertisement delivery;

[0033] The words in the keyword vector that are greater than the key threshold are regarded as demand keywords;

[0034] Adjacent demand keywords are matched through the dictionary in the database, and the matched words are used as demand keyword groups.

[0035] Optionally, the sentence detection network is a trained network; and the first sentence detection network is a network adjusted on the trained sentence detection network.

[0036] Optionally, the training set of the sentence detection network includes training data and labeled data;

[0037] The training data includes a plurality of training sentences; the training sentences include a plurality of training string features; the training string features represent the features of the strings corresponding to the training sentences;

[0038] The labeled data is a vector that labels the positions of words related to advertising in the training sentences as 1, and labels the positions of words unrelated to advertising in the training sentences as 0.

[0039] Optionally, detecting the associated advertisement content according to the fixed interactive sentence length, repeated words and required keywords to obtain the first advertisement label includes:

[0040] Inputting the character string corresponding to the repeated word into the first convolutional network, extracting the features of the repeated word, and obtaining a first advertisement association vector; the first advertisement association vector represents the probability that the repeated word is related to the advertisement keyword; the length of the first advertisement association vector is equal to the number of advertisements to be placed;

[0041] Input the demand keyword into the first convolutional network, extract the features of the demand keyword, and obtain a second advertisement association vector; the second advertisement association vector represents the probability that the repeated word is related to the advertisement keyword; the length of the second advertisement association vector is equal to the number of advertisements to be placed;

[0042] A first advertisement label is obtained based on the first advertisement association vector and the second advertisement association vector.

[0043] Optionally, obtaining the first advertisement label based on the first advertisement association vector and the second advertisement association vector includes:

[0044] Divide each value in the first advertisement association vector by the sentence length to obtain a first advertisement probability vector;

[0045] The values ​​corresponding to the first advertisement probability vector and the second advertisement association vector are averaged to obtain an advertisement association vector;

[0046] The label of the advertisement corresponding to the value greater than other values ​​in the advertisement association vector is used as the first advertisement label.

[0047] In a second aspect, an embodiment of the present invention provides an intelligent message recognition system for advertising AI customer service, including:

[0048] Acquisition module: acquires an interactive sentence and a fixed interactive sentence length and repeated words corresponding to the interactive sentence; the repeated words represent repeated words in the interactive sentence;

[0049] Repeat module: Mark the repeated words according to their positions in the interactive sentences to obtain a repeated marking vector; the repeated marking vector includes n elements; the repeated marking vector marks the repeated words with the same number;

[0050] Sentence detection network module: obtaining a sentence detection network; the sentence detection network is a network capable of extracting keywords related to pre-delivered advertisements in interactive sentences;

[0051] First sentence detection network module: adjusting the sentence convolution network according to the repeated tag vector to obtain a first sentence detection network;

[0052] Demand keyword detection module: according to the interactive sentence, through the first sentence detection network, sentence recognition is performed to obtain demand keywords; the demand keyword keywords represent the target location expected by the user;

[0053] Advertisement message identification module: detects the associated advertisement content according to the fixed interaction sentence length, repeated words and required keywords, and obtains a first advertisement label; the first advertisement label represents the advertisement information represented by the interaction sentence in the chat record.

[0054] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0055] The embodiment of the present invention also provides a message intelligent recognition method and system for AI customer service of advertising delivery, which converts words in a sentence into characters and can increase the features of the words. A first sentence detection network that can control the connection relationship by repeated words is added. For repeated words in different positions in the interactive sentence, it can be known that the features of a word are extracted separately. When the network structure is the same, the corresponding features of the repeated words should be the same, so the features can be fused through the repeated structure, and the fused features can be input into the arranged neurons in the original interactive sentence order through the classification structure to extract the relationship between the words. The user's demand keyword detection is made more accurate. And according to the repeated words in the sentence, according to the composition portrait of the user's interactive sentence, and the association relationship between the demand keyword and the advertising content, the corresponding advertisement corresponding to the advertising content that best suits the user is found together, so as to realize accurate identification of the content information represented by the message in the chat record, and the obtained first advertisement label can accurately represent the advertising information represented by the interactive sentence in the chat record, thereby improving the accuracy of continuing to analyze and identify user needs according to the chat, and based on this, advertising recommendation and delivery can achieve the technical effect of more accurate delivery of advertisements according to user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of a message intelligent identification method for AI customer service of advertising delivery provided by an embodiment of the present invention.

[0057] Figure 2 It is a structural diagram of a first sentence detection network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0059] Example 1

[0060] like Figure 1 As shown, an embodiment of the present invention provides a method for intelligently identifying messages of an AI customer service for advertising, the method comprising:

[0061] S101: Obtain an interactive sentence and a fixed interactive sentence length and repeated words corresponding to the interactive sentence. The repeated words represent repeated words in the interactive sentence.

[0062] In this embodiment, the interactive sentence is in Chinese. The interactive sentence is a sentence composed of multiple words when the user asks the AI ​​customer service robot.

[0063] The length of the interaction sentence is fixed, which is called the fixed interaction sentence length. If the length of the sentence sent by the user is less than the fixed interaction sentence length, the sentence sent by the user is padded with 0 to an interaction sentence of the fixed interaction sentence length. If the length of the sentence sent by the user is greater than the fixed interaction sentence length, the sentence sent by the user is segmented into interaction sentences of the fixed interaction sentence length.

[0064] Among them, in this embodiment, the fixed interactive sentence length is 15, which means a sentence containing 15 words.

[0065] S102: Mark the repeated words according to their positions in the interactive sentences to obtain a repeated marking vector. The repeated marking vector includes n elements. The repeated marking vector marks repeated words as the same.

[0066] The same words in the interactive sentences are marked with the same labels.

[0067] For example, for the question “Is there an electric toothbrush that can brush teeth automatically?”, the repeated label vector is [1,0,1,0,0,0,2,3,0,0,0,3,2].

[0068] S103: Obtain a sentence detection network. The sentence detection network is a network that can extract keywords related to the pre-delivered advertisement in the interactive sentence.

[0069] Wherein, the sentence detection network is a fully connected neural network (Fully Connected Neural Network, FCNN).

[0070] S104: Based on the repeated tag vector, adjust the sentence convolutional network to obtain a first sentence detection network.

[0071] S105: According to the interactive sentence, sentence recognition is performed through the first sentence detection network to obtain a demand keyword. The demand keyword indicates the target location expected by the user.

[0072] S106: Detecting the associated advertisement content according to the fixed interactive sentence length, repeated words and required keywords to obtain a first advertisement label. The first advertisement label represents the advertisement information represented by the interactive sentence in the chat record.

[0073] As an optional implementation manner, the method further includes S107:

[0074] S107: Send a signal for placing the advertisement corresponding to the first advertisement label.

[0075] Optionally, the first sentence detection network adds n storage structures, 1 repetition structure and 1 classification structure to the sentence detection network.

[0076] Wherein, n is a natural number greater than 0, and n is equal to the fixed interaction sentence length. In this embodiment, n is 15.

[0077] Among them, the structural diagram of the first sentence detection network is as follows Figure 2 shown.

[0078] The number of neurons in the input layer of the first sentence detection network is equal to the first interaction length, which is the quotient of the length of the string corresponding to each word matching in the interaction sentence multiplied by the fixed interaction sentence length.

[0079] In this embodiment, for example, "Is there an electric toothbrush that can automatically brush teeth?", the word is converted into a string of length 4. For example, the string corresponding to "yes" is "00x1", the string corresponding to "brush" is "00x2", and the string corresponding to "teeth" is "00x3". If a word in the interactive sentence is fixedly converted into a string of length 4, then the number of neurons in the input layer of the first sentence detection network is 4x15=60.

[0080] A neuron in the first hidden layer of the first sentence neural network is fully connected to m neurons in the input layer corresponding to a word in the interaction sentence.

[0081] Wherein, m is a natural number greater than 0.

[0082] In this embodiment, m is 4.

[0083] Among them, in this embodiment, because the word is converted into a length of 4. One word in the interactive sentence corresponds to 4 neurons in the input layer. For example, according to the order of the interactive sentence "Is there an electric toothbrush that can automatically brush teeth", "no" corresponds to the 5th neuron, 6th neuron, 7th neuron and 8th neuron of the input layer of the first sentence neural network. Therefore, the 5th neuron, 6th neuron, 7th neuron and 8th neuron of the input layer of the first sentence neural network are fully connected to the 2nd neuron of the first hidden layer.

[0084] The neurons of the first hidden layer of the first sentence neural network are connected to a storage structure. The storage structure indicates whether it is a feature corresponding to a repeated word string.

[0085] The n storage structures are fully connected to the 1 repeating structure.

[0086] The repetition structure is connected to a classification structure in which the positions of repeated words are stored.

[0087] The classification structure and the storage structure are fully connected to the neurons of the second hidden layer.

[0088] The second hidden layer is fully connected to two adjacent hidden layers in the plurality of other hidden layers. The other hidden layers are hidden layers other than the first hidden layer and the second hidden layer.

[0089] The last hidden layer is fully connected to the output layer, and the number of neurons in the output layer is equal to the fixed interaction sentence length.

[0090] Optionally, adjusting the sentence convolution network based on the repeated tag vector to obtain a first sentence detection network includes:

[0091] The storage structure includes storage features and storage values. The storage value is initially 0. The storage feature is a feature of a neuron output corresponding to a word in the interactive sentence in the first hidden layer.

[0092] According to the repeat mark vector, the storage value corresponding to the repeat word is set to 1. One element in the repeat mark vector corresponds to one storage structure and one storage value.

[0093] Among them, if the repeated marking vector is [1,0,1,0,0,0,2,3,0,0,0,3,2], the storage value corresponding to the subscript greater than 0 is set to 1. For example, in this embodiment, the length of the fixed interaction sentence is 15, the number of neurons in the first hidden layer is 15, the number of storage structures is 15, and the length of the repeated marking vector is 15. Then the i-th element in the repeated marking vector corresponds to the i-th storage structure, i is greater than or equal to 0, and i is less than 15. Then the storage values ​​of the 1st storage structure, the 3rd storage structure, the 7th storage structure, the 8th storage structure, the 12th storage structure and the 13th storage structure are set to 1.

[0094] If the storage structure is 1, the storage feature is input into the repeating structure.

[0095] In the repetitive structure, the average value of multiple stored features corresponding to the same value in the repetitive tag vector is calculated to obtain a repetitive fusion feature. Multiple labels in the repetitive tag vector correspond to multiple repetitive fusion features.

[0096] Among them, if the repeated label vector is [1,0,1,0,0,0,2,3,0,0,0,3,2], find the storage features corresponding to the same 1 in the repeated label vector, which are the storage features output by the first storage structure and the third storage structure. The storage features output by the first storage structure and the third storage structure are averaged to obtain the repeated fusion feature.

[0097] Wherein, the stored value is used to determine whether the feature output by the neuron corresponding to the first hidden layer is the feature corresponding to the repeated word. The stored value is 0, indicating that the feature output by the neuron corresponding to the first hidden layer is not the feature corresponding to the repeated word, and feature fusion is not performed. The stored value is 1, indicating that the feature output by the neuron corresponding to the first hidden layer is not the feature corresponding to the repeated word, and feature fusion is performed.

[0098] Based on the repeated tag vector and the repeated fusion feature, the classification structure is adjusted to obtain a first sentence detection network.

[0099] Optionally, adjusting the classification structure based on the repeated tag vector and the repeated fusion feature to obtain a first sentence detection network includes:

[0100] According to the order of each word in the interactive sentence, a classification vector is constructed in the classification structure according to the repeated mark vector. The positions corresponding to the same repeated words in the classification vector are set to 1, and the other positions are set to 0. Multiple labels in the repeated mark vector correspond to multiple classification vectors.

[0101] In this embodiment, if the repeated marking vector is [1, 0, 1, 0, 0, 0, 2, 3, 0, 0, 0, 3, 2], the labels are 1, 2, and 3 respectively. Then, three classification vectors are obtained corresponding to the three labels.

[0102] Among them, in this embodiment, the classification vector corresponding to the value of 1 in the repetition mark vector is [1,0,1,0,0,0,0,0,0,0,0,0,0], which means that the repetition fusion feature corresponding to the value of 1 in the repetition mark vector is input into the 1st neuron and the 2nd neuron of the second hidden layer of the first sentence detection network. The classification vector corresponding to the value of 2 in the repetition mark vector is [0,0,0,0,0,0,1,0,0,0,0,0,1], which means that the repetition fusion feature corresponding to the value of 2 in the repetition mark vector is input into the 7th neuron and the 13th neuron of the second hidden layer of the first sentence detection network. The classification vector corresponding to the value of 1 in the repetition mark vector is [0,0,0,0,0,0,0,1,0,0,0,1,0], which means that the repetition fusion feature corresponding to the value of 3 in the repetition mark vector is input into the 8th neuron and the 12th neuron of the second hidden layer of the first sentence detection network.

[0103] According to the corresponding position in the classification vector, the repeated fusion features are input into the corresponding neurons of the second hidden layer.

[0104] If the storage structure is 0, the storage feature is input into the corresponding neuron of the second hidden layer.

[0105] Optionally, performing sentence recognition through a first sentence detection network according to the interactive sentence to obtain the required keywords includes:

[0106] Each word in the interactive sentence is matched with a corresponding string to obtain a string set, wherein a string corresponds to a string in the interactive sentence.

[0107] The character string includes a plurality of characters, and the character string is used to represent a word.

[0108] The database stores a large number of characters and character strings corresponding to the characters. When in use, by searching for each character in the interactive statement, the corresponding character string can be found for matching. In this embodiment, the search algorithm is a sequential search.

[0109] Among them, as in this embodiment, such as "Is there an electric toothbrush that can automatically brush teeth", the character string corresponding to "yes" is "00x1", the character string corresponding to "brush" is "00x2", and the character string corresponding to "teeth" is "00x3".

[0110] Through the above method, using character strings to represent words can facilitate computer calculations, thereby obtaining more information about the words.

[0111] A character string in the character string set is input into the first neural network, and the features of each character in the character string and the features of the connection relationship between the characters are extracted to obtain the first character string feature.

[0112] Among them, the first neural network is a fully connected neural network (Fully Connected Neural Network, FCNN).

[0113] A plurality of first character string features are obtained corresponding to a plurality of character strings in the character string set.

[0114] Input multiple first string features into the first sentence detection network to obtain a keyword vector. The length of the keyword vector is equal to the number of words in the sentence. An element in the keyword vector indicates whether the word in the sentence corresponding to the element is a keyword related to the advertisement delivery.

[0115] The words in the keyword vector that are greater than the key threshold are regarded as demand keywords.

[0116] In this embodiment, the key threshold is 0.9.

[0117] Adjacent demand keywords are matched through the dictionary in the database, and the matched words are used as demand keyword groups.

[0118] Among them, the phrase is used to determine whether the combination of two characters is one word. For example, in this embodiment, two adjacent demand keywords are detected as "tooth" and "brush", and "toothbrush" exists in the phrase, so it can be matched and "toothbrush" is used as the demand keyword phrase.

[0119] Wherein, the dictionary is designed according to the advertisement content and common phrases.

[0120] Optionally, the sentence detection network is a trained network. The first sentence detection network is a network adjusted on the trained sentence detection network.

[0121] Optionally, the training set of the sentence detection network includes training data and labeled data.

[0122] The training data includes a plurality of training sentences. The training sentences include a plurality of training string features. The training string features represent the features of the strings corresponding to the training sentences.

[0123] The labeled data is a vector that labels the positions of words related to advertising in the training sentences as 1, and labels the positions of words unrelated to advertising in the training sentences as 0.

[0124] In this embodiment, for example, "Is there an electric toothbrush that can automatically brush teeth?", "toothbrush" is a keyword related to toothbrush advertisements, and [0,0,0,0,0,0,0,0,0,0,0,1,1,0,0] is used. Because the length of the fixed interactive sentence is 15, the number of vectors output by the sentence detection network is 15, and the 12th and 13th elements in the vector correspond to "toothbrush".

[0125] Optionally, detecting the associated advertisement content according to the fixed interactive sentence length, repeated words and required keywords to obtain the first advertisement label includes:

[0126] The character string corresponding to the repeated word is input into the first convolutional network, and the features of the repeated word are extracted to obtain a first advertisement association vector. The first advertisement association vector includes the degree to which the repeated word is related to the advertisement keyword.

[0127] Among them, the first convolutional network is a fully connected neural network (Fully Connected Neural Network, FCNN).

[0128] The length of the first advertisement association vector is 15.

[0129] The demand keyword is input into the first convolutional network, and the feature of the demand keyword is extracted to obtain a second advertisement association vector. The second advertisement association vector includes the degree to which the demand keyword is related to the advertisement keyword.

[0130] A first advertisement label is obtained based on the first advertisement association vector and the second advertisement association vector.

[0131] Optionally, obtaining the first advertisement label based on the first advertisement association vector and the second advertisement association vector includes:

[0132] Each value in the first advertisement association vector is divided by the sentence length to obtain a first advertisement probability vector.

[0133] The values ​​corresponding to the first advertisement probability vector and the second advertisement association vector are averaged to obtain an advertisement association vector.

[0134] The label of the advertisement corresponding to the value greater than other values ​​in the advertisement association vector is used as the first advertisement label.

[0135] Example 2

[0136] Based on the above-mentioned method for intelligent message recognition of AI customer service for advertising, an embodiment of the present invention also provides a system for intelligent message recognition of AI customer service for advertising, the system comprising an acquisition module, a repetition module, a sentence detection network module, a first sentence detection network module, a demand keyword detection module, an advertising message recognition module and a control delivery module.

[0137] The acquisition module is used to acquire the interactive sentence and the fixed interactive sentence length and repeated words corresponding to the interactive sentence. The repeated words represent repeated words in the interactive sentence.

[0138] The repetition module is used to mark the repeated words according to their positions in the interactive sentences to obtain a repeated marking vector. The repeated marking vector contains n elements. The repeated marking vector marks the repeated words with the same number.

[0139] The sentence detection network module is used to obtain a sentence detection network. The sentence detection network is a network that can extract keywords related to the pre-delivered advertisement in the interactive sentence.

[0140] The first sentence detection network module is used to adjust the sentence convolution network according to the repeated tag vector to obtain the first sentence detection network.

[0141] The demand keyword detection module is used to perform sentence recognition based on the interactive sentence through the first sentence detection network to obtain the demand keyword. The demand keyword indicates the target location expected by the user.

[0142] The advertisement message identification module is used to detect the associated advertisement content according to the fixed interactive sentence length, repeated words and required keywords, and obtain the first advertisement label.

[0143] The control delivery module is used to send a signal for delivering the advertisement corresponding to the first advertisement label.

Claims

1. A method for intelligently identifying messages of AI customer service for advertising, characterized in that: include: Obtaining interactive sentences and fixed interactive sentence lengths and repeated words corresponding to the interactive sentences; The repeated words represent repeated words in the interactive sentence; Marking the repeated words according to their positions in the interactive sentences to obtain a repeated marking vector; the repeated marking vector includes n elements; The repeat mark vector marks repeated words with the same label; Obtaining a sentence detection network; the sentence detection network is a network capable of extracting keywords related to the pre-delivered advertisement in the interactive sentence; Based on the repeated tag vector, adjusting the sentence detection network to obtain a first sentence detection network; According to the interactive sentence, sentence recognition is performed through a first sentence detection network to obtain demand keywords; The demand keyword indicates the target location expected by the user; According to the fixed interactive sentence length, repeated words and required keywords, the associated advertising content is detected to obtain a first advertising label; the first advertising label represents the advertising information represented by the interactive sentence in the chat record; The first sentence detection network adds n storage structures, 1 repetition structure and 1 classification structure to the sentence detection network; A neuron in the first hidden layer of the first sentence detection network is fully connected to m neurons in the input layer corresponding to a word in the interactive sentence; The neurons of the first hidden layer of the first sentence detection network are connected to a storage structure; the storage structure indicates whether it is a feature corresponding to a repeated string; n storage structures are fully connected to 1 repeating structure; The repetition structure is connected to a classification structure; the classification structure stores the positions of repeated words; The classification structure and the storage structure are fully connected with neurons of the second hidden layer; The second hidden layer is fully connected with two adjacent hidden layers in the other hidden layers; The other hidden layers are hidden layers other than the first hidden layer and the second hidden layer; The last hidden layer and output layer are fully connected; The number of neurons in the output layer is equal to the fixed interaction sentence length; The step of adjusting the sentence detection network based on the repeated tag vector to obtain a first sentence detection network includes: The storage structure includes storage features and storage values; the storage value is initially 0; the storage feature is a feature of a neuron output corresponding to a word in the interactive sentence in the first hidden layer; According to the repeat mark vector, the storage value corresponding to the repeat word is set to 1; one element in the repeat mark vector corresponds to one storage structure and one storage value; If the storage structure is 1, the storage feature is input into the repeating structure; In the repetitive structure, the average value of multiple stored features corresponding to the same value in the repetitive tag vector is calculated to obtain a repetitive fusion feature; multiple labels in the repetitive tag vector correspond to multiple repetitive fusion features; Based on the repeated tag vector and the repeated fusion feature, adjusting the classification structure to obtain a first sentence detection network; The step of adjusting the classification structure based on the repeated tag vector and the repeated fusion feature to obtain a first sentence detection network includes: According to the order of each word in the interactive sentence, a classification vector is constructed in the classification structure according to the repeated mark vector; the positions corresponding to the same repeated words in the classification vector are set to 1, and the other positions are set to 0; multiple labels in the repeated mark vector correspond to multiple classification vectors; According to the corresponding position in the classification vector, the repeated fusion feature is input into the corresponding neuron of the second hidden layer; If the storage structure is 0, the storage feature is input into the corresponding neuron of the second hidden layer.

2. The method for intelligently identifying messages of AI customer service for advertising delivery according to claim 1 is characterized in that: The step of performing sentence recognition according to the interactive sentence through a first sentence detection network to obtain the required keywords includes: Match each word in the interactive sentence with a corresponding string to obtain a string set; one string corresponds to one string in the interactive sentence; Obtaining a plurality of first string features corresponding to a plurality of strings in the string set; Inputting a plurality of first character string features into a first sentence detection network to obtain a keyword vector; the length of the keyword vector is equal to the number of words in the sentence; an element in the keyword vector indicates whether the word in the sentence corresponding to the element is a keyword related to advertisement delivery; The words in the keyword vector that are greater than the key threshold are regarded as demand keywords; Adjacent demand keywords are matched by word matching through the dictionary in the database, and the matched words are used as demand keywords.

3. The method for intelligently identifying messages of AI customer service for advertising delivery according to claim 1 is characterized in that: The sentence detection network is a trained network; the first sentence detection network is a network adjusted on the trained sentence detection network.

4. The method for intelligently identifying messages of AI customer service for advertising delivery according to claim 1 is characterized in that: The training set of the sentence detection network includes training data and labeled data; The training data includes multiple training sentences; the training sentences include multiple training string features; The training string feature represents the feature of the string corresponding to the training sentence; The labeled data is a vector that labels the positions of words related to advertising in the training sentences as 1, and labels the positions of words unrelated to advertising in the training sentences as 0.

5. The method for intelligently identifying messages of AI customer service for advertising delivery according to claim 1 is characterized in that: The detecting the associated advertisement content according to the fixed interactive sentence length, repeated words and required keywords to obtain the first advertisement label includes: Inputting the character string corresponding to the repeated word into the first convolutional network, extracting the features of the repeated word, and obtaining a first advertisement association vector; the first advertisement association vector represents the probability that the repeated word is related to the advertisement keyword; the length of the first advertisement association vector is equal to the number of advertisements to be placed; Input the demand keyword into the first convolutional network, extract the features of the demand keyword, and obtain a second advertisement association vector; the second advertisement association vector represents the probability that the repeated word is related to the advertisement keyword; the length of the second advertisement association vector is equal to the number of advertisements to be placed; A first advertisement label is obtained based on the first advertisement association vector and the second advertisement association vector.

6. The method for intelligently identifying messages of AI customer service for advertising delivery according to claim 5 is characterized in that: The obtaining of the first advertisement label based on the first advertisement association vector and the second advertisement association vector includes: Divide each value in the first advertisement association vector by the sentence length to obtain a first advertisement probability vector; The values ​​corresponding to the first advertisement probability vector and the second advertisement association vector are averaged to obtain an advertisement association vector; The label of the advertisement corresponding to the value greater than other values ​​in the advertisement association vector is used as the first advertisement label.

7. An intelligent message recognition system for advertising AI customer service, characterized in that: include: Acquisition module: acquires the interaction sentences and the fixed interaction sentence length and repeated words corresponding to the interaction sentences; The repeated words represent repeated words in the interactive sentence; Repetition module: marking the repeated words according to their positions in the interactive sentences to obtain a repeated marking vector; the repeated marking vector includes n elements; The repeat mark vector marks repeated words with the same label; Sentence detection network module: obtaining a sentence detection network; the sentence detection network is a network capable of extracting keywords related to pre-delivered advertisements in interactive sentences; First sentence detection network module: adjusting the sentence detection network according to the repeated marker vector to obtain a first sentence detection network; Demand keyword detection module: according to the interactive sentence, through the first sentence detection network, sentence recognition is performed to obtain demand keywords; The demand keyword indicates the target location expected by the user; Advertisement message identification module: detects the associated advertisement content according to the fixed interaction sentence length, repeated words and required keywords, and obtains a first advertisement label; the first advertisement label represents the advertisement information represented by the interaction sentence in the chat record; The first sentence detection network adds n storage structures, 1 repetition structure and 1 classification structure to the sentence detection network; A neuron in the first hidden layer of the first sentence detection network is fully connected to m neurons in the input layer corresponding to a word in the interactive sentence; The neurons of the first hidden layer of the first sentence detection network are connected to a storage structure; the storage structure indicates whether it is a feature corresponding to a repeated string; n storage structures are fully connected to 1 repeating structure; The repetition structure is connected to a classification structure; the classification structure stores the positions of repeated words; The classification structure and the storage structure are fully connected with neurons of the second hidden layer; The second hidden layer is fully connected with two adjacent hidden layers in the other hidden layers; The other hidden layers are hidden layers other than the first hidden layer and the second hidden layer; The last hidden layer and output layer are fully connected; The number of neurons in the output layer is equal to the fixed interaction sentence length; The step of adjusting the sentence detection network according to the repeated tag vector to obtain a first sentence detection network includes: The storage structure includes storage features and storage values; the storage value is initially 0; the storage feature is a feature of a neuron output corresponding to a word in the interactive sentence in the first hidden layer; According to the repeat mark vector, the storage value corresponding to the repeat word is set to 1; one element in the repeat mark vector corresponds to one storage structure and one storage value; If the storage structure is 1, the storage feature is input into the repeating structure; In the repetitive structure, the average value of multiple stored features corresponding to the same value in the repetitive tag vector is calculated to obtain a repetitive fusion feature; multiple labels in the repetitive tag vector correspond to multiple repetitive fusion features; Based on the repeated tag vector and the repeated fusion feature, adjusting the classification structure to obtain a first sentence detection network; The step of adjusting the classification structure based on the repeated tag vector and the repeated fusion feature to obtain a first sentence detection network includes: According to the order of each word in the interactive sentence, a classification vector is constructed in the classification structure according to the repeated mark vector; the positions corresponding to the same repeated words in the classification vector are set to 1, and the other positions are set to 0; multiple labels in the repeated mark vector correspond to multiple classification vectors; According to the corresponding position in the classification vector, the repeated fusion feature is input into the corresponding neuron of the second hidden layer; If the storage structure is 0, the storage feature is input into the corresponding neuron of the second hidden layer.

Images