A method and system for selecting aggregated AI answers

Abstract

The present invention discloses a method and system for selecting aggregated AI answers. Whether the answer meets the user's requirements is judged according to the status of the questions asked twice. The repeated keywords and modified keywords of the two questions are used to enhance the accuracy of the answers that meet the user's requirements for the secondary detection. During the secondary detection, Fourier transform and time convolution network are used to obtain the image converted into each word as a feature and the feature of the relationship between multiple words. And the features corresponding to each word are extracted through a three-dimensional convolution kernel. First, the features of the keywords that the user still needs to know are detected by repeating the keywords, and then the answers corresponding to the keywords that were detected incorrectly by the computer when the question was asked for the first time are detected by modifying the keywords. Thereby achieving a more accurate technical effect of selecting answers.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a method and system for selecting aggregated AI answers. Background Art

[0002] At present, AI question-answering mainly detects and identifies based on big data. Although a lot of data is stored in the knowledge base, it is still difficult to understand the questions raised by users and compare and match them. Because usually based on the questions raised by users, multiple answers related to the questions raised by users can be found, but how to design and train the network so that the network can intelligently detect the user's answers is still a problem. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for selecting aggregated AI answers 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 selecting an aggregated AI answer, including:

[0005] A knowledge database, a first user question statement and a corresponding first keyword set are obtained; the first user question statement indicates a statement in which the user asks a question for the first time; the first keyword set includes a plurality of first keywords; the first keyword indicates a keyword in the first user question statement; the knowledge database is a database storing a plurality of data for finding answers;

[0006] Obtaining a first detection answer based on the first user question statement and the knowledge database through an answer detection network;

[0007] Obtaining a second user question statement and a corresponding second keyword set; the second user question statement represents a statement in which the user continues to ask questions after outputting the first detection answer; the second keyword set includes a plurality of second keywords; the second keyword represents a keyword in the second user question statement;

[0008] Based on the second keyword set and the first keyword set, a repeated keyword and a modified keyword are obtained; the modified keyword is a keyword in the second user's question statement that is different from that in the first user's question statement; the repeated keyword is a keyword in the second user's question statement that is the same as that in the first user's question statement;

[0009] Through the repeated association network, based on the repeated keywords, the corresponding answers in the knowledge database are found to obtain repeated sentences and repeated answer sentence segments;

[0010] Using a local detection network, based on the modification keyword and the repeated sentence, the type corresponding to the repeated sentence is detected to obtain the repeated sentence type;

[0011] The repeated answer sentence segment corresponding to the repeated sentence type is sent as a selected answer.

[0012] Optionally, the step of finding a corresponding answer in a knowledge database based on the repeated keyword through a repeated association network to obtain a repeated answer sentence segment includes:

[0013] Based on the repeated keywords, find data of the same page containing the repeated keywords in the knowledge database to obtain the repeated keyword page and the corresponding multiple repeated statements;

[0014] Through the repeated association network, the association between two adjacent repeated sentences is detected to obtain the repeated sentence association value;

[0015] If the repetition sentence association value is greater than the repetition association threshold, the portion between two repetition sentences in the repetition keyword page is used as a repetition answer sentence segment.

[0016] Optionally, the local detection network includes a local feature extraction network and a category detection network;

[0017] Obtain training data and annotated data; the annotated data includes an annotation modification keyword and an annotation detection category; the annotation modification keyword represents a keyword in the second keyword set that is different from the first keyword in the first keyword set; the training data is a repeated sentence; the annotation detection category represents the meaning represented by the repeated sentence containing the modification keyword;

[0018] Based on the annotated first modification keywords and the training data, a local feature extraction network is trained using the first modification network;

[0019] The training data is input into the local detection network, and the loss is calculated with the labeled detection category. The local feature extraction network and the category detection network trained by the labeled modification keywords are trained to obtain a trained local detection network.

[0020] Optionally, detecting the correlation between two adjacent repeated sentences through a repeated correlation network to obtain a repeated sentence correlation value includes:

[0021] Obtaining a first repeated statement and a second repeated statement; the first repeated statement and the second repeated statement are adjacent repeated statements;

[0022] Performing Fourier transformation on the first repeated sentence to obtain a first repeated sentence transformation spectrum; performing Fourier transformation on the second repeated sentence to obtain a second repeated sentence transformation spectrum;

[0023] Inputting the first repeated sentence into the temporal convolutional network to obtain a first convolutional feature map; inputting the second repeated sentence into the temporal convolutional network to obtain a second convolutional feature map;

[0024] The first convolution feature map and the first repeated sentence transformation spectrum map are fused to obtain a first fusion map; the second convolution feature map and the second repeated sentence transformation spectrum map are fused to obtain a second fusion map;

[0025] Based on the first fusion graph and the second fusion graph, association judgment is performed through a repeated association network to obtain repeated sentence association values.

[0026] Optionally, performing association judgment based on the first fusion graph and the second fusion graph through a repeated association network to obtain a repeated sentence association value includes:

[0027] Obtain a three-dimensional convolution kernel; the number of channels of the three-dimensional convolution kernel is equal to the number of channels of the first fusion image; the width of the three-dimensional convolution kernel is equal to the width of the first fusion image; the length of the three-dimensional convolution kernel is 2;

[0028] On the first fusion graph, the three-dimensional convolution kernel performs convolution from top to bottom with a step size of 1, detects features corresponding to each word in the first repeated sentence and joint features between words, and obtains a first fusion graph;

[0029] On the second fusion graph, the three-dimensional convolution kernel performs convolution from top to bottom with a step size of 1, detects features corresponding to each word in the second repeated sentence and joint features between words, and obtains a second fusion graph;

[0030] The first fusion graph and the second fusion graph are input into an associative neural network to determine the associative relationship between two repeated sentences and obtain an associative value of the repeated sentences.

[0031] Optionally, the step of training a local feature extraction network using the first modification network based on the labeled first modification keyword and the training data includes:

[0032] Inputting the training data into a local feature extraction network to obtain a first prediction feature;

[0033] Inputting the first prediction feature into a first modification network to obtain a predicted first modification keyword;

[0034] The predicted first modification keyword and the labeled first modification keyword are used to obtain losses, and the first modification network and the local feature extraction network are trained to obtain a trained local feature extraction network.

[0035] Optionally, based on the repeated keyword, finding data of the same page containing the repeated keyword in the knowledge database to obtain the repeated keyword page and the corresponding multiple repeated statements includes:

[0036] Detecting repeated keywords in a knowledge database, finding a page containing the repeated keywords, and obtaining a repeated keyword page; the repeated keyword page contains a plurality of repeated keywords;

[0037] On the repeated keyword page, the context content of the repeated keyword is detected to obtain a repeated sentence; and multiple repeated sentences are obtained corresponding to multiple repeated keywords on the repeated keyword page.

[0038] Optionally, obtaining repeated keywords and modified keywords based on the second keyword set and the first keyword set includes:

[0039] Matching the second keyword set with the first keyword set to obtain a matching value; the matching value indicates the overlap ratio of the first keyword in the first keyword set and the second keyword in the second keyword set;

[0040] If the matching value is less than the matching threshold, inputting the second user question statement into the answer detection network;

[0041] If the matching value is greater than or equal to the matching threshold, find the second keyword in the second keyword set that is the same as the first keyword set as the repeated keyword; find the second keyword in the second keyword set that is different from the first keyword set as the modified keyword.

[0042] Optionally, if the matching value is less than a matching threshold, the answer detection network is trained using the first user question statement and the first detection answer.

[0043] In a second aspect, an embodiment of the present invention provides a selection system for aggregating AI answers, including:

[0044] Acquisition module: obtain a knowledge database, a first user question statement and a corresponding first keyword set; the first user question statement represents the statement in which the user asks a question for the first time; the first keyword set includes multiple first keywords; the first keyword represents the keyword in the first user question statement; the knowledge database is a database storing multiple data for finding answers; obtain a second user question statement and a corresponding second keyword set; the second user question statement represents the statement in which the user continues to ask questions after outputting the first detection answer; the second keyword set includes multiple second keywords; the second keyword represents the keyword in the second user question statement;

[0045] Answer detection module: obtaining a first detection answer based on the first user's question statement and the knowledge database through an answer detection network;

[0046] Keyword judgment module: based on the second keyword set and the first keyword set, obtain repeated keywords and modified keywords; the modified keywords are keywords in the second user's question statement that are different from those in the first user's question statement; the repeated keywords are keywords in the second user's question statement that are the same as those in the first user's question statement;

[0047] Repeat answer module: through the repeated association network, based on the repeated keywords, find the corresponding answer in the knowledge database to obtain the repeated sentence and the repeated answer sentence segment;

[0048] Local detection module: detects the type corresponding to the repeated sentence based on the modification keyword and the repeated sentence through the local detection network, and obtains the type of the repeated sentence;

[0049] Sending module: sending the repeated answer sentence segment corresponding to the repeated sentence type as a selected answer.

[0050] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0051] An embodiment of the present invention also provides a method and system for selecting an aggregated AI answer, the method comprising: obtaining a knowledge database, a first user question statement and a corresponding first keyword set; the first user question statement represents the statement in which the user asks a question for the first time; the first keyword set includes multiple first keywords; the first keyword represents the keyword in the first user question statement; the knowledge database is a database storing multiple data for finding answers; through an answer detection network, a first detection answer is obtained based on the first user question statement and the knowledge database; a second user question statement and a corresponding second keyword set are obtained; the second user question statement represents the statement in which the user continues to ask questions after the first detection answer is output; the second keyword set includes A plurality of second keywords; the second keywords represent keywords in the second user's question statement; based on the second keyword set and the first keyword set, repeated keywords and modified keywords are obtained; the modified keywords are keywords in the second user's question statement that are different from those in the first user's question statement; the repeated keywords are keywords in the second user's question statement that are the same as those in the first user's question statement; through a repeated association network, based on the repeated keywords, corresponding answers are found in a knowledge database to obtain repeated sentences and repeated answer sentence segments; through a local detection network, based on the modified keywords and repeated sentences, the types corresponding to the repeated sentences are detected to obtain repeated sentence types; and the repeated answer sentence segments corresponding to the repeated sentence types are sent as selected answers.

[0052] In the present invention, whether the answer meets the user's requirements is judged according to the status of the questions asked twice. Thus, the repeated keywords and modified keywords of the two questions are used to enhance the accuracy of the answers that meet the user's requirements for the secondary detection. During the secondary detection, Fourier transform and time convolution network are used to obtain the image converted into each word as a feature and the feature of the relationship between multiple words. And the features corresponding to each word are extracted through a three-dimensional convolution kernel. First, the features of the keywords that the user still needs to know are detected by repeating the keywords, and then the answers corresponding to the keywords that were detected incorrectly by the computer when the question was asked for the first time are detected by modifying the keywords. Thereby achieving a more accurate technical effect of selecting answers. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a method for selecting aggregated AI answers provided by an embodiment of the present invention.

[0054] Figure 2 It is a structural diagram of a local detection network in a method for selecting an aggregated AI answer provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0056] Example 1

[0057] like Figure 1 As shown, an embodiment of the present invention provides a method for selecting an aggregated AI answer, the method comprising:

[0058] S101: Obtain a knowledge database, a first user question statement and a corresponding first keyword set; the first user question statement indicates a statement in which the user asks a question for the first time; the first keyword set includes a plurality of first keywords; the first keyword indicates a keyword in the first user question statement; the knowledge database is a database storing a plurality of data for finding answers;

[0059] The knowledge database is data crawled by a crawler.

[0060] The keywords are classified according to parts of speech. For example, in this embodiment, nouns and interrogative prepositions are used as keywords, and the words corresponding to the nouns and interrogative prepositions in the first user's question sentence are added to the first keyword set.

[0061] S102: Obtain a first detection answer through an answer detection network based on the first user question statement and a knowledge database.

[0062] The answer detection network is a fully connected neural network (FCN), and the answer detection network is trained by using numbers marked with answers.

[0063] S103: Obtain a second user question statement and a corresponding second keyword set; the second user question statement represents a statement in which the user continues to ask questions after outputting the first detection answer; the second keyword set includes multiple second keywords; the second keyword represents a keyword in the second user question statement.

[0064] The second user question sentence is a sentence obtained through the same data receiving interface as the first user question sentence.

[0065] S104: Based on the second keyword set and the first keyword set, repeated keywords and modified keywords are obtained; the modified keywords are keywords in the second user's question statement that are different from those in the first user's question statement; the repeated keywords are keywords in the second user's question statement that are the same as those in the first user's question statement.

[0066] S105: Find the corresponding answer in the knowledge database based on the repeated keywords through the repeated association network to obtain repeated sentences and repeated answer sentence segments.

[0067] The repeated answer sentence segment represents multiple sentences. The repeated answer sentence segment contains repeated sentences.

[0068] S106: Using a local detection network, based on the modification keyword and the repeated sentence, detect the type corresponding to the repeated sentence to obtain the repeated sentence type.

[0069] The structural diagram of the local detection network is as follows: Figure 2 shown.

[0070] In this embodiment, the repeated sentence type indicates that the detected sentence belongs to the type of content that is expected to obtain a structure. For example, if the question is "What is the structure of a neural network", the repeated sentence type is the "content" type. If the question is "What is the category of a neural network", the repeated sentence type is the "classification" type.

[0071] S107: Send the repeated answer sentence segment corresponding to the repeated sentence type as a selected answer.

[0072] Optionally, the step of finding a corresponding answer in a knowledge database based on the repeated keyword through a repeated association network to obtain a repeated answer sentence segment includes:

[0073] Based on the repeated keywords, data of the same page containing the repeated keywords are found in the knowledge database to obtain a repeated keyword page and corresponding multiple repeated statements.

[0074] Among them, because the data in the knowledge database is the data crawled by the crawler, the crawled pages are recorded during the crawling, and the data of the same page are judged.

[0075] Through the repeated association network, the association between two adjacent repeated sentences is detected to obtain the repeated sentence association value.

[0076] The range of the repeated sentence association value is 0-1, and a larger value of the repeated sentence association value indicates that the contents represented by two adjacent repeated sentences are more identical.

[0077] If the repetition sentence association value is greater than the repetition association threshold, the portion between two repetition sentences in the repetition keyword page is used as a repetition answer sentence segment.

[0078] Among them, because the repeated sentence association value is greater than the repeated association threshold, it means that the two adjacent repeated sentences express the same content, so the content between the repeated sentences in the two vectors is also describing the same content, or is associated with it, so the part between the two repeated sentences is taken as the repeated answer sentence segment.

[0079] Optionally, the local detection network includes a local feature extraction network and a category detection network.

[0080] Among them, in this embodiment, the local feature extraction network is a convolutional neural network (CNN); the category detection network is a fully connected neural network (FCN).

[0081] Obtain training data and annotation data; the annotation data includes annotation modification keywords and annotation detection categories; the annotation modification keywords represent keywords in a second keyword set that are different from first keywords in a first keyword set; the training data are historical repeated sentences used for training; the annotation detection category represents the meaning represented by the repeated sentences containing the modification keywords.

[0082] Based on the labeled first modification keywords and training data, the first modification network is used to train the local feature extraction network.

[0083] The training data is input into the local detection network, and the loss is calculated with the labeled detection category. The local feature extraction network and the category detection network trained by the labeled modification keywords are trained to obtain a trained local detection network.

[0084] Among them, the cross entropy loss function is used to calculate the loss.

[0085] Optionally, detecting the correlation between two adjacent repeated sentences through a repeated correlation network to obtain a repeated sentence correlation value includes:

[0086] Obtaining a first repeated statement and a second repeated statement; the first repeated statement and the second repeated statement are adjacent repeated statements;

[0087] The first repeated sentence is subjected to Fourier transformation to obtain a first repeated sentence transformation spectrum graph; the second repeated sentence is subjected to Fourier transformation to obtain a second repeated sentence transformation spectrum graph.

[0088] The first repetitive sentence transformation spectrum and the second repetitive sentence transformation spectrum have the same size; the width of the first repetitive sentence transformation spectrum is equal to the length of the set input sentence. In this embodiment, the length of the input sentence is set to 32 words, that is, the width of the first repetitive sentence transformation spectrum is 32.

[0089] The first repeated sentence is input into the temporal convolutional network to obtain a first convolutional feature map; the second repeated sentence is input into the temporal convolutional network to obtain a second convolutional feature map.

[0090] The first convolutional feature map and the second convolutional feature map have the same size.

[0091] In this embodiment, each word in the first repeated sentence is sequentially input into a temporal convolutional network (TCN) according to the word order in the first repeated sentence for feature extraction.

[0092] The first convolution feature map and the first repeated sentence transformation spectrum map are fused to obtain a first fusion map; the second convolution feature map and the second repeated sentence transformation spectrum map are fused to obtain a second fusion map.

[0093] Among them, the width of the first convolution feature map is equal to the width of the first repeated sentence transformation spectrum map, and the height of the first convolution feature map is equal to the height of the first repeated sentence transformation spectrum map.

[0094] The first convolution feature map and the first repeated sentence transformation spectrogram are superimposed in the channel direction to obtain a first fusion map. If the number of channels of the first convolution feature map is n, the number of channels of the first repeated sentence transformation spectrogram is m, the number of channels of the first fusion map is n+m, and n and m are natural numbers greater than 0.

[0095] Based on the first fusion graph and the second fusion graph, association judgment is performed through a repeated association network to obtain repeated sentence association values.

[0096] Optionally, performing association judgment based on the first fusion graph and the second fusion graph through a repeated association network to obtain a repeated sentence association value includes:

[0097] Obtain a three-dimensional convolution kernel; the number of channels of the three-dimensional convolution kernel is equal to the number of channels of the first fusion image; the width of the three-dimensional convolution kernel is equal to the width of the first fusion image; the length of the three-dimensional convolution kernel is 2;

[0098] On the first fusion graph, the three-dimensional convolution kernel performs convolution from top to bottom with a step size of 1 to detect features corresponding to each word in the first repeated sentence and joint features between words to obtain a first fusion graph.

[0099] The number of values ​​in the first fusion graph is equal to the number of characters in the first repeated sentence.

[0100] Among them, after Fourier transform, each horizontal axis corresponds to the features of different characters; the features extracted by the temporal convolutional network are the fused features of multiple characters, so convolution is performed from top to bottom through the three-dimensional convolution kernel, and a similar convolution method is used for the features of each character to obtain the first fusion graph that can determine each character.

[0101] On the second fusion graph, the three-dimensional convolution kernel performs convolution from top to bottom with a step size of 1, detects features corresponding to each word in the second repeated sentence and joint features between words, and obtains a second fusion graph;

[0102] Inputting the first fusion graph and the second fusion graph into an association neural network to determine the association relationship between two repeated sentences and obtain a repeated sentence association value;

[0103] The associative neural network is a discriminator network. During the training process, if the sentences corresponding to the first fusion graph and the second fusion graph respectively express the same meaning, the association value of the marked repeated sentences is 1. If they are different, the association value of the marked repeated sentences is 0.

[0104] Optionally, the step of training a local feature extraction network using the first modification network based on the labeled first modification keyword and the training data includes:

[0105] The training data is input into a local feature extraction network to extract features of repeated sentences and corresponding modified keywords to obtain predicted local features.

[0106] The predicted local features are input into a first modification network, and keywords in the first keyword set that are different from the second user's question statement are detected to obtain predicted first modification keywords.

[0107] The predicted first modification keyword and the labeled first modification keyword are used to obtain losses, and the first modification network and the local feature extraction network are trained to obtain a trained local feature extraction network.

[0108] The trained local feature extraction network is capable of extracting features for detecting keywords.

[0109] Among them, the loss is calculated through the cross entropy loss function.

[0110] Optionally, based on the repeated keyword, finding data of the same page containing the repeated keyword in the knowledge database to obtain the repeated keyword page and the corresponding multiple repeated statements includes:

[0111] Detect repeated keywords in a knowledge database, find pages containing repeated keywords, and obtain repeated keyword pages; the repeated keyword pages contain multiple repeated keywords.

[0112] Among them, because the data in the knowledge database is the data crawled by the crawler, the crawled pages are recorded during the crawling, and the same page is marked; through the sequential search algorithm, the positions of multiple repeated keywords in a page are found.

[0113] On the repeated keyword page, the context content of the repeated keyword is detected to obtain a repeated sentence; and multiple repeated sentences are obtained corresponding to multiple repeated keywords on the repeated keyword page.

[0114] The content of the context of the repeated keyword is a sentence containing the repeated keyword separated by a period as a repeated sentence.

[0115] Optionally, obtaining repeated keywords and modified keywords based on the second keyword set and the first keyword set includes:

[0116] The second keyword set is matched with the first keyword set to obtain a matching value; the matching value indicates the overlap ratio of the first keyword in the first keyword set and the second keyword in the second keyword set.

[0117] If the matching value is less than the matching threshold, the second user question statement is input into the answer detection network.

[0118] The matching threshold is 0.85.

[0119] If the matching value is greater than the matching threshold, find the second keyword in the second keyword set that is the same as the first keyword set as the repeated keyword; find the second keyword in the second keyword set that is different from the first keyword set as the modified keyword.

[0120] Optionally, if the matching value is less than a matching threshold, the answer detection network is trained using the first user question statement and the first detection answer.

[0121] The answer detection network is a fully connected neural network (FCN), and the answer detection network is trained by using numbers marked with answers.

[0122] Example 2

[0123] Based on the above-mentioned method for selecting aggregated AI answers, an embodiment of the present invention also provides a system for selecting aggregated AI answers, the system comprising an acquisition module, an answer detection module, a keyword judgment module, a repeated answer module, a local detection module and a sending module.

[0124] The acquisition module is used to obtain a knowledge database, a first user question statement and a corresponding first keyword set; the first user question statement represents the statement in which the user asks a question for the first time; the first keyword set includes multiple first keywords; the first keyword represents the keyword in the first user question statement; the knowledge database is a database that stores multiple data for finding answers; obtain a second user question statement and a corresponding second keyword set; the second user question statement represents the statement in which the user continues to ask questions after outputting the first detection answer; the second keyword set includes multiple second keywords; the second keyword represents the keyword in the second user question statement.

[0125] The answer detection module is used to obtain a first detection answer based on the first user question statement and the knowledge database through an answer detection network.

[0126] The keyword judgment module is used to obtain repeated keywords and modified keywords based on the second keyword set and the first keyword set; the modified keywords are keywords in the second user's question statement that are different from those in the first user's question statement; the repeated keywords are keywords in the second user's question statement that are the same as those in the first user's question statement.

[0127] The repeated answer module is used to find the corresponding answer in the knowledge database based on the repeated keywords through the repeated association network, and obtain the repeated sentences and repeated answer sentence segments.

[0128] The local detection module is used to detect the type corresponding to the repeated statement based on the modification keyword and the repeated statement through the local detection network to obtain the repeated statement type.

[0129] The sending module is used to send the repeated answer sentence segment corresponding to the repeated sentence type as a selected answer.

Claims

1. A method for selecting aggregated AI answers, characterized in that: include: Obtaining a knowledge database, a first user question statement, and a corresponding first keyword set; The first user question statement refers to the statement in which the user asks a question for the first time; the first keyword set includes a plurality of first keywords; the first keyword refers to the keyword in the first user question statement; The knowledge database is a database storing a plurality of data for finding answers; Obtaining a first detection answer based on the first user question statement and the knowledge database through an answer detection network; Obtaining a second user question statement and a corresponding second keyword set; The second user question statement refers to a statement in which the user continues to ask questions after the first detection answer is output; the second keyword set includes a plurality of second keywords; the second keyword refers to a keyword in the second user question statement; Based on the second keyword set and the first keyword set, a repeated keyword and a modified keyword are obtained; the modified keyword is a keyword in the second user's question statement that is different from that in the first user's question statement; The repeated keyword is the same keyword in the second user's question statement as in the first user's question statement; Through the repeated association network, based on the repeated keywords, the corresponding answers in the knowledge database are found to obtain repeated sentences and repeated answer sentence segments; Using a local detection network, based on the modification keyword and the repeated sentence, the type corresponding to the repeated sentence is detected to obtain the repeated sentence type; Sending the repeated answer sentence segment corresponding to the repeated sentence type as a selected answer; The method of finding the corresponding answer in the knowledge database based on the repeated keyword through the repeated association network to obtain the repeated answer sentence segment includes: Based on the repeated keywords, find data of the same page containing the repeated keywords in the knowledge database to obtain the repeated keyword page and the corresponding multiple repeated statements; Through the repeated association network, the association between two adjacent repeated sentences is detected to obtain the repeated sentence association value; If the repetition sentence association value is greater than the repetition association threshold, the portion between two repetition sentences in the repetition keyword page is used as a repetition answer sentence segment.

2. A method for selecting aggregated AI answers according to claim 1, characterized in that: The local detection network includes a local feature extraction network and a category detection network; Obtain training data and annotated data; the annotated data includes an annotation modification keyword and an annotation detection category; the annotation modification keyword represents a keyword in the second keyword set that is different from a first keyword in the first keyword set; The training data is a historical repeated sentence used for training; the label detection category indicates the meaning of the repeated sentence containing the modified keyword; Based on the annotated first modification keywords and the training data, a local feature extraction network is trained using the first modification network; The training data is input into the local detection network, and the loss is calculated with the labeled detection category. The local feature extraction network and the category detection network trained by the labeled modification keywords are trained to obtain a trained local detection network.

3. The method for selecting an aggregated AI answer according to claim 1, characterized in that: The method of detecting the correlation between two adjacent repeated sentences through the repeated correlation network to obtain the repeated sentence correlation value includes: Obtaining a first repeated statement and a second repeated statement; the first repeated statement and the second repeated statement are adjacent repeated statements; Performing Fourier transformation on the first repeated sentence to obtain a first repeated sentence transformation spectrum; performing Fourier transformation on the second repeated sentence to obtain a second repeated sentence transformation spectrum; Inputting the first repeated sentence into the temporal convolutional network to obtain a first convolutional feature map; inputting the second repeated sentence into the temporal convolutional network to obtain a second convolutional feature map; The first convolution feature map and the first repeated sentence transformation spectrum map are fused to obtain a first fusion map; the second convolution feature map and the second repeated sentence transformation spectrum map are fused to obtain a second fusion map; Based on the first fusion graph and the second fusion graph, association judgment is performed through a repeated association network to obtain repeated sentence association values.

4. A method for selecting aggregated AI answers according to claim 3, characterized in that: The performing association judgment based on the first fusion graph and the second fusion graph through the repeated association network to obtain the repeated sentence association value includes: Obtain a three-dimensional convolution kernel; the number of channels of the three-dimensional convolution kernel is equal to the number of channels of the first fusion image; the width of the three-dimensional convolution kernel is equal to the width of the first fusion image; the length of the three-dimensional convolution kernel is 2; On the first fusion graph, the three-dimensional convolution kernel performs convolution from top to bottom with a step size of 1, detects features corresponding to each word in the first repeated sentence and joint features between words, and obtains a first fusion graph; On the second fusion graph, the three-dimensional convolution kernel performs convolution from top to bottom with a step size of 1, detects features corresponding to each word in the second repeated sentence and joint features between words, and obtains a second fusion graph; The first fusion graph and the second fusion graph are input into an associative neural network to determine the associative relationship between two repeated sentences and obtain an associative value of the repeated sentences.

5. The method for selecting an aggregated AI answer according to claim 2, characterized in that: The method of training a local feature extraction network using a first modification network based on annotating the first modification keyword and the training data includes: Inputting the training data into a local feature extraction network, extracting features of repeated sentences and corresponding modified keywords, and obtaining predicted local features; Inputting the predicted local features into a first modification network, detecting keywords in the first keyword set that are different from the second user's question sentence, and obtaining predicted first modification keywords; The predicted first modification keyword and the labeled first modification keyword are used to obtain losses, and the first modification network and the local feature extraction network are trained to obtain a trained local feature extraction network.

6. The method for selecting an aggregated AI answer according to claim 1, characterized in that: Based on the repeated keyword, data of the same page containing the repeated keyword is found in the knowledge database to obtain the repeated keyword page and the corresponding multiple repeated statements, including: Detecting repeated keywords in a knowledge database, finding a page containing the repeated keywords, and obtaining a repeated keyword page; the repeated keyword page contains a plurality of repeated keywords; On the repeated keyword page, the context content of the repeated keyword is detected to obtain a repeated sentence; and multiple repeated keywords in the repeated keyword page correspond to multiple repeated sentences.

7. The method for selecting an aggregated AI answer according to claim 1, characterized in that: The obtaining of repeated keywords and modified keywords based on the second keyword set and the first keyword set includes: Matching the second keyword set with the first keyword set to obtain a matching value; the matching value indicates the overlap ratio of the first keyword in the first keyword set and the second keyword in the second keyword set; If the matching value is less than the matching threshold, inputting the second user question statement into the answer detection network; If the matching value is greater than or equal to the matching threshold, find the second keyword in the second keyword set that is the same as the first keyword set as the repeated keyword; find the second keyword in the second keyword set that is different from the first keyword set as the modified keyword.

8. The method for selecting an aggregated AI answer according to claim 7, characterized in that: If the matching value is less than the matching threshold, the answer detection network is trained using the first user question statement and the first detection answer.

9. A selection system for aggregating AI answers, characterized in that: include: Acquisition module: obtains the knowledge database, the first user question statement and the corresponding first keyword set; The first user question statement refers to the statement in which the user asks a question for the first time; the first keyword set includes a plurality of first keywords; the first keyword refers to the keyword in the first user question statement; The knowledge database is a database storing a plurality of data for finding answers; Obtaining a second user question statement and a corresponding second keyword set; The second user question statement refers to a statement in which the user continues to ask questions after the first detection answer is output; the second keyword set includes a plurality of second keywords; the second keyword refers to a keyword in the second user question statement; Answer detection module: obtaining a first detection answer based on the first user's question statement and the knowledge database through an answer detection network; Keyword judgment module: based on the second keyword set and the first keyword set, obtains repeated keywords and modified keywords; the modified keywords are keywords in the second user's question statement that are different from those in the first user's question statement; The repeated keyword is the same keyword in the second user's question statement as in the first user's question statement; Repeat answer module: through the repeated association network, based on the repeated keywords, find the corresponding answer in the knowledge database to obtain the repeated sentence and the repeated answer sentence segment; Local detection module: detects the type corresponding to the repeated sentence based on the modification keyword and the repeated sentence through the local detection network, and obtains the type of the repeated sentence; A sending module: sending the repeated answer sentence segment corresponding to the repeated sentence type as a selected answer; The method of finding the corresponding answer in the knowledge database based on the repeated keyword through the repeated association network to obtain the repeated answer sentence segment includes: Based on the repeated keywords, find data of the same page containing the repeated keywords in the knowledge database to obtain the repeated keyword page and the corresponding multiple repeated statements; Through the repeated association network, the association between two adjacent repeated sentences is detected to obtain the repeated sentence association value; If the repetition sentence association value is greater than the repetition association threshold, the portion between two repetition sentences in the repetition keyword page is used as a repetition answer sentence segment.

Images