Data processing method, device and equipment

Abstract

The invention discloses a data processing method, device and equipment, and the method comprises the steps: obtaining first problem data inputted by a user in a current session; based on the dialogue data of the current session and the dialogue data of multiple rounds of the historical sessions, determining a target round of the target historical session, the dialogue data of each round including question data and reply data; based on a target round of the target historical session, obtaining target data corresponding to the target round, the target data being data obtained in a process of processing question data of the target round to obtain reply data of the target round; and processing the first question data by using the target data, and determining first reply data corresponding to the first question data.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a data processing method, device and equipment. Background Art

[0002] Currently, some devices have a dialogue function, where the device can receive questions input by the user and then output the answers obtained after processing the questions to the user.

[0003] In the related art, in order to obtain a more accurate answer, when receiving a question, the device will process the current question and the previous rounds of conversations together, and then make a reply based on the other conversation content, where each round of conversation includes the input question and the processed answer.

[0004] The problem with this processing method is that each time a new question is processed, it is necessary to process the new question and the content of multiple rounds of conversations in the past, resulting in a large amount of data to be processed each time a question is processed, a long time to obtain an answer, and low response efficiency. Summary of the invention

[0005] To this end, this application discloses the following technical solutions:

[0006] The first aspect of the present application provides a data processing method, comprising:

[0007] Get the first question data entered by the user in the current session;

[0008] Determine a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations, wherein the conversation data of each round includes a question data and an answer data;

[0009] Based on the target round of the target historical conversation, obtaining target data corresponding to the target round, wherein the target data is data obtained in the process of processing the question data of the target round to obtain the answer data of the target round;

[0010] The first question data is processed using the target data to determine first answer data corresponding to the first question data.

[0011] Optionally, obtaining the conversation data of the current session includes:

[0012] The conversation data before the first question data in the current conversation, or the conversation data of the historical conversation that meets specific conditions, is obtained as the conversation data of the current conversation, wherein the conversation data of the current conversation includes N rounds of conversations, where N is greater than or equal to 1.

[0013] Optionally, determining a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations includes:

[0014] Compare the N rounds of conversation data in the current session with multiple rounds of conversation data in the historical session. If there are N rounds of conversation data in the historical session that match the N rounds of conversation data in the current session, determine the historical session as the target historical session, and obtain the target round of the target historical session, where N is greater than or equal to 1.

[0015] Optionally, determining a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations includes:

[0016] Based on the first question data of the current session, the conversation data of the current session, and the conversation data of multiple rounds of historical sessions, a target round of the target historical session is determined.

[0017] Optionally, determining a target round of a target historical conversation based on the first question data of the current conversation, the conversation data of the current conversation, and the conversation data of multiple rounds of historical conversations includes:

[0018] In the case where second question data matching the first question data is found in the historical session, a round corresponding to the second question data is used as a target round, and target data of the target round is obtained, wherein the target data of the round corresponding to the second question data is data obtained in the process of obtaining second answer data from the second question data;

[0019] When no second question data matching the first question data is found in the historical conversation, the conversation data in the current conversation is compared with conversation data of multiple rounds in the historical conversation to determine a target round of the target historical conversation.

[0020] Optionally, also include:

[0021] Determine the hash value of each round of dialogue in the session, and / or the hash value of the question data in each round, wherein the hash value is used to determine the comparison between the dialogue data of the current session and the dialogue data of multiple rounds of historical sessions, and / or the comparison between the question data in the current session and the question data of multiple rounds of historical sessions.

[0022] Optionally, also include:

[0023] In a case where the first question data is question data of the first round of the current conversation, processing the first question data and determining first answer data corresponding to the first question data;

[0024] Alternatively, the first question data is processed using the conversation data of the historical conversation that meets specific conditions to determine the first answer data corresponding to the first question data.

[0025] Optionally, also include:

[0026] Get the subject of the current session;

[0027] A historical session whose corresponding topics in the plurality of historical sessions match the topic of the current session is determined as a target historical session.

[0028] Optionally, also include:

[0029] Obtain predicted question data based on the conversation data of the current conversation, the first question data and the first answer data;

[0030] Processing the predicted question data to obtain predicted answer data;

[0031] In the case where the obtained third question data is the same as the predicted question data, the predicted answer data is used as the third answer data corresponding to the third question data, and the third question data is the question data obtained after the first question data.

[0032] A second aspect of the present application provides a data processing device, including:

[0033] An input unit, used to obtain first question data input by a user in a current session;

[0034] a determination unit, configured to determine a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations, wherein the conversation data of each round includes a question data and an answer data;

[0035] An acquisition unit, configured to acquire target data corresponding to the target round based on the target round of the target historical conversation, wherein the target data is data obtained in the process of processing the question data of the target round to obtain the answer data of the target round;

[0036] A processing unit is used to process the first question data using the target data to determine first answer data corresponding to the first question data.

[0037] A third aspect of the present application provides an electronic device, including an application module and at least one model;

[0038] The application module is used to receive data input by a user, and determine a target model to be called based on analyzing the data input by the user. The application module can call the at least one model;

[0039] The target model is used to obtain data input by a user, determine output data corresponding to the data input by the user based on the user's historical data and the target data corresponding to the historical data, and return the output data to the application module so that the application module outputs the output data to the user, wherein the historical data includes input data input by the user and output data corresponding to the input data, and the target data corresponding to the historical data is data obtained in the process of processing the input data to obtain the output data. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0041] Figure 1 is a flow chart of a data processing method provided in an embodiment of the present application;

[0042] Figure 2 is a schematic diagram of a session tree provided in an embodiment of the present application;

[0043] Figure 3 is a flow chart of a method for determining a target historical session and a target round provided by an embodiment of the present application;

[0044] Figure 4 It is a schematic diagram of a hash value of a conversation provided in an embodiment of the present application;

[0045] Figure 5 It is a schematic diagram of deleting duplicate nodes in a session tree of a current session provided by an embodiment of the present application;

[0046] Figure 6 It is a structural diagram of a data processing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0048] This application embodiment provides a data processing method, see Figure 1 , is a flow chart of the method, and the method may include the following steps.

[0049] S101, obtaining first question data input by a user in a current session.

[0050] The data processing method provided in this embodiment can be executed by any electronic device with a conversation function.

[0051] The first question data may be the question data input by the user in the first round of the current session, that is, the first question data input by the user in the current session.

[0052] The first question data may be question data input by the user in the N+1th round of the current conversation, where N is greater than or equal to 1. For example, the first question data may be question data input by the user in the 4th round of the current conversation. In the current conversation, the user and the electronic device of this embodiment have had 3 rounds of conversations before obtaining the first question data.

[0053] The first question data may include multiple forms, for example, at least one of voice data, image data, and text data, without limitation.

[0054] As an example, the first question data may be text-formatted question data “Please recommend a computer suitable for playing games”.

[0055] S102, determining a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations, wherein the conversation data of each round includes a question data and an answer data.

[0056] A historical conversation may include several conversations that occurred before the current conversation. At the end of each conversation, the device may save all questions and answers obtained during the conversation as a historical conversation. Each historical conversation may include several rounds of conversation data.

[0057] For example, during a conversation, the user inputs question 1, and the device processes question 1 and outputs corresponding answer 1. Then the user inputs question 2, and the device processes question 2 and inputs corresponding answer 2. The conversation ends at this time. The device can save the above questions and answers as a historical conversation, where question 1 and answer 1 constitute the first round of conversation data in the historical conversation, and question 2 and answer 2 constitute the second round of conversation data in the historical conversation.

[0058] A device may determine that a session has ended based on a number of methods.

[0059] First, after the device outputs a reply, if no new input is obtained within a preset waiting time, it can be determined that the current session is over.

[0060] Second, after outputting the reply, the device may determine that the current session is ended in response to an operation by the user to end the current session.

[0061] Third, after outputting a reply, the device determines that the number of rounds contained in the current session has reached the upper limit, then determines that the current session has ended. For example, the upper limit is that a session contains a maximum of 20 rounds of conversation data. Then, after the device outputs the reply to the 20th round of the current session, it determines that the number of rounds has reached the upper limit.

[0062] After a conversation is determined to be over in any of the above three ways, if the device obtains a new input question, the question can be determined as the first round of questions in a new conversation process, thereby starting a new conversation process.

[0063] Fourth, when the device obtains a newly input question, if it determines that the question has no semantic relevance to the questions and answers of the previous round, the device can determine that the question is the first round of questions in a new conversation, and determine that the conversation to which the questions and answers of the previous round belong has ended.

[0064] For example, after the device outputs a corresponding answer to a question about purchasing electronic products, it receives a question about route navigation. The device determines that the conversation to which the previously obtained question and the corresponding answer belong has ended, opens a new conversation, and uses the currently obtained question about route navigation as the first round of questions in the currently opened conversation.

[0065] In step S102, a target historical session may be determined in the historical sessions, and then a target round of the target historical session may be determined. The target historical session may be a historical session whose session content is related to the current session, and the target round may be a round in the target historical session that is related to the conversation data and the first question data.

[0066] One way to determine the target historical session may be to obtain the subject of the current session; and determine a historical session whose corresponding subject in multiple historical sessions matches the subject of the current session as the target historical session.

[0067] Each time a conversation begins, the device can determine a number of prompt words that represent the subject of the conversation based on the questions and answers obtained in the conversation. When determining the target historical conversation, the prompt words of the current conversation can be compared with the prompt words of the historical conversation. If the prompt words of the current conversation are the same as the prompt words of a certain historical conversation, or the prompt words of a certain historical conversation include the prompt words of the current conversation, it can be determined that the subject of the historical conversation matches the subject of the current conversation, and the historical conversation can be determined as the target historical conversation.

[0068] The conversation data of the current session may include a variety of conversation data related to the first question data of the current session.

[0069] S103, based on the target round of the target historical conversation, obtaining target data corresponding to the target round, where the target data is data obtained in the process of processing the question data of the target round to obtain the answer data of the target round.

[0070] Each time the device obtains a question data, it can use various processing models capable of processing question data to process the question data to obtain corresponding answer data. The processing model used can be a large language model or other models, without limitation.

[0071] In this embodiment, each time the device processes a round of problem data using the processing model, the data obtained during the processing can be saved as the target data of the round.

[0072] Exemplarily, if the processing model used is a large language model, the device can save the KV Cache data obtained when processing a round of problem data as the target data of this round. The KV Cache data of a round is related to the conversation data of this round, the conversation data of multiple rounds before this round, and other reference data that may be used when processing the problem of this round.

[0073] Exemplarily, the target round may be the K+1th round of the target historical conversation. When the device processes the question data of the K+1th round, the conversation data of the first K rounds of the target historical conversation and the question data of the K+1th round may be input into the large language model, and a portion of the reply data of the K+1th round obtained during the processing will also be introduced into the KV Cache data. Therefore, the KV Cache data of the K+1th round obtained during the processing is related to the conversation data of the first K+1 rounds.

[0074] S104: Process the first question data using the target data to determine first answer data corresponding to the first question data.

[0075] In step S104, the target data of the target round obtained and the first question data of S101 may be input into a processing model, and the first answer data corresponding to the first question data may be obtained by processing the processing model.

[0076] Optionally, if the corresponding target round is not determined in S102, the first question data may be processed using the dialogue data of the current session to determine the first answer data corresponding to the first question data.

[0077] The beneficial effects of this embodiment are:

[0078] On the one hand, the target data obtained when processing the question data of the target round is related to the conversations in the target round and the rounds before it, while the conversation data of the target round is related to the first question data currently obtained. Therefore, by using the target data of the target round to process the first question data, the context of the question can be considered when answering the question, which is conducive to obtaining a more accurate answer.

[0079] On the other hand, when obtaining the newly input first question data, the method of this embodiment only needs to use the target data obtained in the process of processing the question data of the target round to process the first question data, so as to determine the corresponding first reply data. Compared with the solution of processing the content of multiple rounds of conversations in the previous technology in the related art, this embodiment only needs to process the data of a specific target round, and thus can significantly reduce the amount of data that needs to be processed, thereby shortening the time for processing problems and improving processing efficiency.

[0080] Therefore, the method of this embodiment can improve the efficiency of answering questions to a certain extent without reducing the accuracy of the answers.

[0081] In some embodiments, when the first question data is the question of the N+1th round of the current session, and the conversation data of the current session includes conversation data of N rounds (N is greater than or equal to 1), the conversation data of the current session can be obtained as follows:

[0082] First, the conversation data before the first question data in the current conversation can be obtained as the conversation data of the current conversation;

[0083] Second, the conversation data of the historical conversations that meet specific conditions can be obtained as the conversation data of the current conversation.

[0084] In the first acquisition method, the device can use all rounds of conversation data before the first question data in the current session as the conversation data of the current session. For example, if the first question data is the question of the N+1th round of the current session, then the conversation data of the current session can include the conversation data of the first N rounds.

[0085] In the second acquisition method, if a historical conversation that ended before the current conversation (referred to as the previous historical conversation) meets specific conditions, the conversation data of all rounds contained in the historical conversation can be obtained as the conversation data of the current round, or the conversation data of multiple rounds in the historical conversation that are temporally closer to the first question data can be obtained as the conversation data of the current round.

[0086] The specific conditions may include a time continuity condition. If the device outputs the answer to the last round in the previous historical conversation and obtains the first round of questions in the current conversation within a preset interval (e.g., 1 minute, 2 minutes, not limited to the specific one), it can be determined that the previous historical conversation meets the time continuity condition. If the interval between the answer to the last round in the previous historical conversation and the first round of questions in the current conversation exceeds the interval, it can be determined that the time continuity condition is not met.

[0087] The specific conditions may include semantically relevant conditions. After each conversation starts, the device may determine one or more prompt words corresponding to the conversation based on the conversation data of each round obtained during the conversation. These prompt words may represent the subject of the conversation. For example, the prompt words of a conversation may include "medical treatment", "cold", etc. If the prompt words of the previous historical conversation are the same as the prompt words of the current conversation, it can be determined that the previous historical conversation meets the semantically relevant conditions. If the two are different, it can be determined that the previous historical conversation does not meet the semantically relevant conditions.

[0088] The specific conditions may also include a time continuity condition and a semantic relevance condition, that is, the conversation data of the previous historical conversation that satisfies both the time continuity condition and the semantic relevance condition may be used as the conversation data of the current conversation.

[0089] Optionally, if there is a previous historical conversation that meets specific conditions, the conversation data of the previous historical conversation and the conversation data of N rounds before the first question data in the current session can be used as the conversation data of the current session. If there is no previous historical conversation that meets the specific conditions, the conversation data of the current session can be obtained only in the first way.

[0090] In some optional embodiments, based on the conversation data of the current conversation and the conversation data of multiple rounds of the historical conversation, the target round of the target historical conversation may be determined by:

[0091] Compare the N rounds of conversation data in the current session with multiple rounds of conversation data in the historical session. If there are N rounds of conversation data in the historical session that match the N rounds of conversation data in the current session, determine the historical session as the target historical session, and obtain the target round of the target historical session, where N is greater than or equal to 1.

[0092] In the above embodiment, the conversation data of each round in the conversation data of the current conversation obtained in S102 may be compared with the conversation data of multiple rounds contained in a plurality of historical conversations to determine the target historical conversation.

[0093] In this embodiment, in order to facilitate comparison, the device may save each historical session in the form of a session tree.

[0094] by Figure 2 (1) is an example. When starting conversation 1, the device can determine the prompt word of the conversation based on the obtained question, and save the prompt word as the root node of conversation 1. When processing the question data of the first round of conversation 1 and outputting the answer data of the first round, the device can save the conversation data of the first round (including the questions and answers of the first round) as node 1 corresponding to the first round, and connect it to the root node of conversation 1.

[0095] When processing the second round of questions in conversation 1 and outputting the second round of replies, the device may save the first round of conversation data and the second round of conversation data as node 2 corresponding to the second round, and connect them after node 1 of conversation 1; when processing the third round of questions in conversation 1 and outputting the third round of replies, the device may save the first three rounds of conversation data as node 3 corresponding to the third round, and connect them after node 2 of conversation 1. After outputting the third round of replies, the device determines that conversation 1 ends, thereby obtaining the following result: Figure 2 The session tree of session 1 shown in (1).

[0096] That is to say, for each conversation, the device can determine a node for each round of the conversation. The node corresponding to each round can save the conversation data of this round and the previous round in the conversation. The node corresponding to each round serves as a child node of the node corresponding to the previous round, thereby forming a conversation tree for the conversation.

[0097] Furthermore, the device may merge the same nodes in the session trees of different sessions and use the different nodes as branches, thereby forming a session tree with several branches.

[0098] by Figure 2 As an example, the device obtains the session tree corresponding to session 2 in the aforementioned manner, and finds that the conversation data of the first and second rounds of session 2 are the same as the conversation data of the first and second rounds of session 1, and the conversation data of the third round of session 2 is different from the conversation data of the third round of session 1. Therefore, the nodes corresponding to the first and second rounds in the session data of session 2 can be deleted, and the node corresponding to the third round of session 2 can be connected to node 2 as node 4 to form Figure 2 The conversation tree with a branch structure shown in (2) has nodes 3 and 4 as child nodes of node 2. Nodes 3 and 4 reflect two different conversations that may occur after the conversation at node 2 occurs.

[0099] In the case of branches, a session tree can correspond to multiple different historical sessions, for example Figure 2 (2) can correspond to two historical sessions, session 1 and session 2.

[0100] Based on the above session tree structure, see Figure 3 , the method of comparing N rounds of conversation data in the current conversation with multiple rounds of conversation data in the historical conversation may include:

[0101] 1. Select a session tree and compare the prompt word of the current session with the root node of the session tree;

[0102] 2. If the prompt word of the current session does not match the root node of the session tree, select another session tree and return to step 1;

[0103] 3. If the prompt word of the current session matches the root node of the session tree, compare the conversation data of the kth round in the conversation data of the current session with the node of the kth layer of the session tree;

[0104] When step 3 is executed for the first time, the value of k is 1, and the root node in the session tree can be used as the 0th layer, the child nodes of the 0th layer as the 1st layer, the child nodes of the 1st layer as the 2nd layer, and so on;

[0105] In some optional embodiments, the root node of the session tree may also correspond to the first round of dialogue in the session, in which case the root node may be used as the first layer;

[0106] 4. If the conversation data of the kth round in the conversation data of the current conversation does not match any node of the kth layer of the conversation tree, select another conversation tree and return to step 1;

[0107] 5. If the conversation data of the kth round in the conversation data of the current conversation matches any node of the kth layer of the conversation tree, and k is not equal to N, increase k by 1 and return to step 3;

[0108] 6. If the conversation data of the kth round in the conversation data of the current conversation matches any node of the kth layer of the conversation tree, and k is equal to N, a historical conversation corresponding to the conversation tree is determined to be the target historical conversation.

[0109] Each time step 3 is executed, the conversation data of the kth round can be compared with those nodes in the kth layer whose parent nodes match the k-1th round, and there is no need to compare those nodes whose parent nodes do not match the k-1th round.

[0110] by Figure 2 (2) is an example, assuming Figure 2 In (2), both node 3 and node 4 have several child nodes. The conversation in the third round of the current session matches node 4 but does not match node 3. When comparing the conversation in the fourth round of the current session with the fourth-level nodes of the session tree, it is only necessary to compare the conversation in the fourth round with the child nodes of node 4, and it is not necessary to compare the conversation in the fourth round with the child nodes of node 4.

[0111] In step 6, the node in the Nth layer of the session tree that is successfully matched may be obtained, and the historical session corresponding to the node may be determined as the target historical session.

[0112] by Figure 2 (2) is an example. Assuming N is 3, after comparison, it is found that the conversation data of the third round of the current conversation matches the node 4 of the third layer, so the conversation 2 corresponding to the node 4 is determined to be the target historical conversation.

[0113] The above comparison method is equivalent to, starting from the first round of the conversation data of the current conversation, comparing each round of conversation data with the nodes of the corresponding layer in the conversation tree corresponding to the historical conversation one by one. If there is a node in each layer of the conversation tree that can be matched successfully, then the historical conversation corresponding to the successfully matched node in the conversation tree is determined as the target historical conversation.

[0114] After the target historical session is determined, the target round of the target historical session can be determined as follows:

[0115] A successfully matched node in the Nth layer is obtained, and the round corresponding to the node is determined as the target round.

[0116] by Figure 2 (2) is an example. Assume that N is 3, and the third round of the conversation data of the current conversation successfully matches node 4. Therefore, conversation 2 corresponding to node 4 is obtained as the target historical conversation, and the third round of conversation 2 corresponding to node 4 is taken as the target round.

[0117] In the above embodiments, comparing the conversation data of the current session with the node means comparing the conversation data of the current session with the conversation data of the round corresponding to the node. For example, comparing the third round of conversation of the current session with node 4 means comparing the third round of conversation of the current session with the third round of conversation of session 2 corresponding to node 4.

[0118] The advantages of establishing a session tree and performing comparison based on the session tree in the above manner are:

[0119] On the one hand, after merging multiple historical conversations into a conversation tree with branches, the number of repeated comparisons can be reduced. On the other hand, based on the branch structure in the conversation tree, some unnecessary comparison content can be omitted each time the comparison is performed. For example, in the aforementioned example, only the child nodes of node 4 need to be compared, and the child nodes of node 3 do not need to be compared. Therefore, comparing based on the conversation tree in the above manner can improve the efficiency of determining the target historical conversation and the target round.

[0120] In some embodiments, each node of the conversation tree may also only store the conversation data of a corresponding round, without storing the conversation data of the previous round. Figure 2 In (2), node 2 only stores the second round of conversation data of session 1, node 3 only stores the third round of conversation data of session 1, and node 4 only stores the third round of conversation data of session 2.

[0121] In some embodiments, a conversation tree may not be established. In this case, for each historical conversation, the prompt words of the historical conversation can be compared with the prompt words of the current conversation. If the match fails, another historical conversation is compared. If the match succeeds, the conversation data of the first N rounds of the historical conversation and the first N rounds of the current conversation are compared in sequence, that is, the first round of the historical conversation is compared with the first round of the current conversation, the first two rounds of the historical conversation are compared with the first two rounds of the current conversation, and so on.

[0122] If all N rounds are matched successfully, it is determined that the historical session is the target historical session. If a match fails when comparing any round, another historical session is compared.

[0123] After the target historical session is determined, the Nth round of the target historical session is determined as the target round.

[0124] In some embodiments, the method of comparing N rounds of conversation data in the current conversation with multiple rounds of conversation data in the historical conversation may also be:

[0125] For each historical session, the first round of conversation data in the conversation data of the current session is compared with each round of conversation data of the historical session. If the conversation data of any round in the historical session successfully matches the first round of conversation data in the current session, the round in the historical session that successfully matches is determined to be a candidate round. Then, starting from the candidate round, the second to Nth rounds of the current session can be compared with each round after the candidate round in the historical session in sequence, that is, the second round of the current session is compared with the first round after the candidate round in the historical session, the third round of the current session is compared with the second round after the candidate round in the historical session, and so on. If a round fails to match, continue to compare another historical session. If N consecutive rounds including the candidate round in the historical session and the first N rounds of conversation data of the current session are all successfully matched, then the historical session is determined to be the target historical session, and the last round of the N consecutive rounds that successfully matched in the target historical session is determined as the target round.

[0126] Exemplarily, assuming that the first question data is the question of the 4th round of the current session, N is equal to 3, and the 1st round of the current session matches the 3rd round of the historical session 2 successfully, then continue to compare the 2nd round of the current session with the 4th round of the historical session 2, and compare the 3rd round of the current session with the 5th round of the historical session 2. After comparison, it is found that 3 consecutive rounds 3, 4 and 5 in the historical session 2 are successfully matched with the first 3 rounds of the current session. Therefore, the historical session 2 can be determined as the target historical session, and the 5th round of the historical session 2 can be determined as the target round.

[0127] The above comparison method is also applicable to the case of session trees. For each session tree, the first round of the current session can be compared with each node in the session tree except the root node. If it matches successfully with any node, the node is determined to be a candidate node. Then, the candidate node is used as the first layer, and the nodes of each subsequent layer are compared one by one with the conversation data of each round of the current session, that is, the nodes of the second layer are compared with the second round of the current session, and the nodes of the third layer are compared with the third round of the current session. If there are matching nodes in each layer, the target historical session can be determined based on the last node that successfully matches the Nth round of the current session, and the last round contained in the node is determined as the target round.

[0128] In some embodiments, in order to further improve the efficiency of determining the target historical conversation and the target historical round, the first question data can be referred to when determining the target round. That is, the target round of the target historical conversation can be determined based on the first question data of the current conversation, the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations.

[0129] Optionally, when introducing the first question data and determining the target round of the target historical session, the first question data may be compared with each question data contained in each historical session to search for second question data matching the first question data from the historical session.

[0130] In the case where second question data matching the first question data is found in the historical session, the historical session to which the second question data belongs is determined as the target historical session, the round corresponding to the second question data is used as the target round, and the target data of the target round is obtained, where the target data of the round corresponding to the second question data is the data obtained in the process of obtaining the second answer data from the second question data;

[0131] When the second question data matching the first question data is not found in the historical conversation, the conversation data in the current conversation is compared with the conversation data of multiple rounds in the historical conversation to determine the target round of the target historical conversation.

[0132] In the latter case, the manner of comparing the conversation data in the current conversation with the conversation data of multiple rounds in the historical conversations can be referred to in the aforementioned embodiment and will not be described in detail.

[0133] Optionally, when second question data that matches the first question data is found in the historical session, the conversation data of the first N rounds of the second question data in the historical session can be compared with the conversation data of the first N rounds of the current session one by one. If the conversation data of any round fails to match, it is determined that the round corresponding to the second question data is not the target round, and the target historical session and target round are determined in the manner of comparing the conversation data in the current session with the conversation data of multiple rounds in the historical session in the aforementioned embodiment. If the conversation data of each round matches successfully, the historical session to which the second question data belongs can be determined as the target historical session, and the round corresponding to the second question data can be used as the target round.

[0134] After finding the second question data that matches successfully, continue to compare the first N rounds of the second question data with the first N rounds of the current session. Its functions are:

[0135] In some cases, due to the difference in contextual contents of the first question data and the second question data, the first question data and the second question data may be matched successfully but have different semantics.

[0136] For example, the first question data and the second question data are both "recommend a computer suitable for playing games", and the two can be matched successfully, but the first few rounds of conversations in the first question data are about computer products newly launched in the past year, and the first few rounds of conversations in the second question data are about computer products of Brand A. Therefore, the semantics of the first question data is to recommend a computer suitable for playing games from computer products newly launched in the past year, and the semantics of the second question data is to recommend a computer suitable for playing games from computer products of Brand A. The two have different semantics. In this case, if the target data of the corresponding round of the second question data is used to process the first question data, an incorrect answer is likely to be obtained.

[0137] By comparing the first N rounds of the second question data with the first N rounds of the current conversation, it is possible to determine whether the contexts of the first question data and the second question data match, thereby determining whether the semantics of the first question data and the second question data match through whether the contexts match, thereby avoiding processing the first question data with the target data of the corresponding round of the second question data when the semantics of the two are different, thereby achieving the effect of reducing the error rate of the response.

[0138] Optionally, when second question data matching the first question data is found, and the first N rounds of the second question data and the first N rounds of the current session all match successfully, the second reply data corresponding to the second question data can also be directly obtained and the second reply data can be directly output as the first reply data.

[0139] When the first question data and the second question data match successfully, and the first N rounds of the second question data and the first N rounds of the current session all match successfully, it can be considered that the first question data and the second question data are two identical questions. Therefore, the first question data is no longer processed, and the second answer data obtained when processing the second question data can be directly used as an answer to the first question data. In this way, the efficiency of answering repeated questions can be improved.

[0140] Optionally, after the first question data is introduced, the N rounds of conversation data in the current conversation can be compared with the conversation data of multiple rounds of historical conversations according to the method of the above embodiment to determine the target historical conversation (for example, according to Figure 3 method for comparison).

[0141] After the target historical session is determined, the question data of the N+1th round of the target historical session can be compared with the first question data. If the question data of the N+1th round and the first question data match successfully, the question data of the N+1th round of the target historical session can be determined to be the second question data, and the N+1th round of the target historical session can be determined to be the target round. If the question data of the N+1th round and the first question data fail to match, the Nth round of the target historical session can be determined to be the target round.

[0142] In some optional embodiments, after completing each round of dialogue, the device may:

[0143] Determine the hash value of each round of dialogue in the session, and / or the hash value of the question data in each round, the hash value is used to determine the comparison between the dialogue data of the current session and the dialogue data of multiple rounds of historical sessions, and / or the comparison between the question data in the current session and the question data of multiple rounds of historical sessions.

[0144] Among them, the hash value of a round of conversation can be obtained directly by processing a question data and an answer data of the round using any hash algorithm in the relevant technical field, or by processing the session data of the round and all previous rounds in the corresponding session using a hash algorithm.

[0145] For the latter case, Figure 4For example, at the end of the first round of conversation in session 1, the conversation data of the first round can be processed to obtain the hash value of the first round. At the end of the second round of conversation in session 1, the conversation data of the second round and the previous conversation data can be processed, that is, the conversation data of the first round and the second round can be processed to obtain the hash value of the second round. At the end of the third round of conversation in session 1, the conversation data from the first round to the third round can be processed to obtain the hash value of the third round, and so on.

[0146] Based on the hash value of each round of conversation data, the conversation data of a round in the current session and the conversation data of a round in the historical session may be compared by comparing the hash values ​​of the two. If the hash values ​​of the two are the same, it is determined that the two are matched successfully; if the hash values ​​of the two are different, it is determined that the two are matched unsuccessfully.

[0147] Alternatively, the similarity of the hash values ​​of the two may be determined. If the similarity is greater than a threshold, the match is successful; if the similarity is less than or equal to the threshold, the match fails.

[0148] The hash value of the question data in each round can be obtained by processing the question data of the round using a hash algorithm. When comparing the first question data with the question data of each round in the historical session, the hash value of the first question data can be compared with the hash value of the question data of each round in the historical session. If the hash value of the first question data is the same as the hash value of the question data in the historical session, or the similarity is greater than a threshold, it can be determined that the first question data and the question data in the historical session match successfully. If the hash value of the first question data is not the same as the hash value of the question data in the historical session, or the similarity is less than or equal to the threshold, it can be determined that the first question data and the question data in the historical session fail to match.

[0149] Optionally, when the first question data is question data of the first round of the current session, the first question data may be processed in at least one of the following ways to obtain corresponding first answer data:

[0150] 1. Processing the first question data to determine the first answer data corresponding to the first question data;

[0151] Second, using the conversation data of the historical conversation that meets the specific conditions, the first question data is processed to determine the first answer data corresponding to the first question data.

[0152] In processing method 1, the device can directly input the first question data into a processing model for processing question data, and use the processing model to process the first question data to obtain first answer data.

[0153] In processing method two, the device can determine a historical conversation that meets the aforementioned time continuity condition and / or semantic relevance condition, and input all the conversation data and the first question data contained in the historical conversation into the processing model, and the processing model processes all the conversation data and the first question data contained in the historical conversation to obtain the first reply data, or, input several conversation data contained in the historical conversation that are closer in time to the first question data and the first question data into the processing model, and the processing model processes them to obtain the first reply data.

[0154] Optionally, after outputting the first reply data, the following steps may also be performed:

[0155] Obtain predicted question data based on the conversation data of the current session, the first question data and the first answer data;

[0156] Processing the predicted question data to obtain predicted answer data;

[0157] When the obtained third question data and the predicted question data are the same, the predicted answer data is used as the third answer data corresponding to the third question data, and the third question data is the question data obtained after the first question data.

[0158] In this embodiment, the device can input the conversation data of the current session, the first question data and the first answer data into a processing model after outputting the first answer data and before obtaining new question data, and use the data input into the processing model to predict new questions that subsequent users may ask, thereby obtaining predicted question data output by the processing model, and then process the predicted question data to obtain predicted answer data corresponding to the predicted question data.

[0159] The method of processing the prediction question data may be consistent with the method of processing the first question data in the aforementioned embodiment, and will not be described in detail.

[0160] When the device obtains a new question, that is, obtains the aforementioned third question data, it can compare the third question data and the predicted question data. If the two match successfully, the predicted answer data can be directly output as the third answer data corresponding to the third question data without the need to process the third answer data again. If the two fail to match, the third question data can be processed in the same way as the aforementioned processing of the first question data to obtain the third answer data.

[0161] In some cases, after outputting the first reply data, the user generally spends a certain amount of time browsing and processing the first reply data and will not immediately input new questions. The method of this embodiment can use the time the user spends browsing and processing the first reply data to predict the questions that the user may ask later and respond to the predicted questions in advance. In this way, when the question actually input by the user matches the predicted question, the reply obtained in advance can be output immediately, further improving the efficiency of the conversation.

[0162] In some embodiments, when determining that the current session has ended, the session tree of the current session can also be formed using conversation data from multiple rounds in the current session in the manner described above to determine the session tree. After obtaining the session tree of the current session, the nodes of the session tree of the current session can be compared with the nodes of the session tree corresponding to the historical session.

[0163] If it is found that some nodes in the session tree of the current session are repeated with those in the session tree corresponding to the historical session, those repeated nodes in the session tree of the current session can be deleted, and the non-duplicate nodes are used as branches of the session tree corresponding to the historical session and connected to the session tree corresponding to the historical session.

[0164] by Figure 5 For example, suppose Figure 5 (1) is the session tree corresponding to the historical session. Figure 5 (2) is the session tree of the current session. After comparison, it is found that the root node of the session tree of the current session is the same as the root node of the session tree corresponding to the historical session. Node 5 of the session tree of the current session is the same as node 1 of the session tree corresponding to the historical session. The other nodes are different. Therefore, the root node and node 5 of the session tree of the current session can be deleted, and the remaining nodes 6 and 7 are connected to node 1 as branches to form Figure 5 The session tree shown in (3).

[0165] Optionally, the session tree of the current session can be synchronously established during the session of the current session, that is, the root node is determined at the beginning of the current session, and then each time a round of dialogue in the current session is completed, a node corresponding to the round of dialogue is established in the session tree of the current session, thereby the session tree of the current session can be immediately obtained when the current session ends.

[0166] Optionally, in the case of directly saving the conversation data, the conversation data of each round in the current session can be compared with the conversation data of each historical session at the end of the current session. If the conversation data of a round in the current session is repeated with the conversation data of a round in a historical session, then only this round of the current session can be recorded, and the conversation data of this round is recorded as repeated with the conversation data of the corresponding round in the historical session. There is no need to repeatedly save the conversation data of this round in the current session to save storage space.

[0167] Exemplarily, assuming that when the current conversation ends, it is determined that the current conversation contains three rounds of conversation, wherein the second round of conversation repeats the second round of conversation in historical conversation 1, that is, the two contain the same questions and the same answers. Then, when recording the current conversation, it is possible to record the first round of the current conversation and its conversation data, the third round of the current conversation and its conversation data, the second round of the current conversation, and record that the second round of the conversation is the same as the second round of historical conversation 1, thereby saving storage space for recording the conversation data of the second round.

[0168] In some embodiments, the start time of a historical session can be recorded each time the data of the historical session is saved. On this basis, expired historical sessions whose duration from the start time to the current time exceeds a certain threshold (for example, 7 days, 30 days, etc.) can be regularly searched for, and the conversation data corresponding to the expired historical sessions can be deleted to save storage space.

[0169] This embodiment also provides a data processing device, see Figure 6 , is a schematic diagram of the device, and the device may include the following units.

[0170] An input unit 601 is used to obtain first question data input by a user in a current session;

[0171] A determination unit 602, configured to determine a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations, wherein the conversation data of each round includes a question data and an answer data;

[0172] An acquisition unit 603 is used to acquire target data corresponding to the target round based on the target round of the target historical conversation, where the target data is data obtained in the process of processing the question data of the target round to obtain the answer data of the target round;

[0173] The processing unit 604 is configured to process the first question data using the target data to determine first answer data corresponding to the first question data.

[0174] Optionally, when the determining unit 602 obtains the dialog data of the current session, it can be used to:

[0175] The conversation data before the first question data in the current conversation, or the conversation data of the historical conversation that meets specific conditions, is obtained as the conversation data of the current conversation, wherein the conversation data of the current conversation includes N rounds of conversations, where N is greater than or equal to 1.

[0176] Optionally, when the determination unit 602 determines the target round of the target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of the historical conversation, it can be used to:

[0177] Compare the N rounds of conversation data in the current session with multiple rounds of conversation data in the historical session. If there are N rounds of conversation data in the historical session that match the N rounds of conversation data in the current session, determine the historical session as the target historical session, and obtain the target round of the target historical session, where N is greater than or equal to 1.

[0178] Optionally, when the determination unit 602 determines the target round of the target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of the historical conversation, it can be used to:

[0179] Based on the first question data of the current session, the conversation data of the current session, and the conversation data of multiple rounds of historical sessions, a target round of the target historical session is determined.

[0180] Optionally, when the determination unit 602 determines the target round of the target historical conversation based on the first question data of the current conversation, the conversation data of the current conversation, and the conversation data of multiple rounds of the historical conversation, it can be used to:

[0181] When second question data matching the first question data is found in the historical session, the round corresponding to the second question data is used as the target round, and target data of the target round is obtained. The target data of the round corresponding to the second question data is the data obtained in the process of obtaining the second answer data from the second question data.

[0182] When the second question data matching the first question data is not found in the historical conversation, the conversation data in the current conversation is compared with the conversation data of multiple rounds in the historical conversation to determine the target round of the target historical conversation.

[0183] Optionally, the determining unit 602 is further configured to:

[0184] Determine the hash value of each round of dialogue in the session, and / or the hash value of the question data in each round, the hash value is used to determine the comparison between the dialogue data of the current session and the dialogue data of multiple rounds of historical sessions, and / or the comparison between the question data in the current session and the question data of multiple rounds of historical sessions.

[0185] Optionally, the processing unit 604 is further configured to:

[0186] In the case where the first question data is question data of the first round of the current conversation, processing the first question data and determining first answer data corresponding to the first question data;

[0187] Alternatively, the first question data is processed using the conversation data of the historical conversation that meets the specific conditions to determine the first answer data corresponding to the first question data.

[0188] Optionally, the determining unit 602 is further configured to:

[0189] Get the topic of the current session;

[0190] A historical session whose corresponding topics in the multiple historical sessions match the topic of the current session is determined as a target historical session.

[0191] Optionally, the processing unit 604 is further configured to:

[0192] Obtain predicted question data based on the conversation data of the current session, the first question data and the first answer data;

[0193] Processing the predicted question data to obtain predicted answer data;

[0194] When the obtained third question data and the predicted question data are the same, the predicted answer data is used as the third answer data corresponding to the third question data, and the third question data is the question data obtained after the first question data.

[0195] The working principle of the data processing device of this embodiment can refer to the relevant steps of the data processing method provided in any embodiment, and will not be described in detail.

[0196] This embodiment also provides an electronic device for implementing the method of the above embodiment, and the electronic device may include an application module and at least one model;

[0197] An application module, used for receiving data input by a user, and determining a target model to be called based on analyzing the data input by the user. The application module can call at least one model;

[0198] The target model is used to obtain the data input by the user, determine the output data corresponding to the data input by the user based on the user's historical data and the target data corresponding to the historical data, and return the output data to the application module so that the application module outputs the output data to the user, wherein the historical data includes the input data input by the user and the output data corresponding to the input data, and the target data corresponding to the historical data is the data obtained in the process of processing the input data to obtain the output data.

[0199] The electronic device of this embodiment can be any electronic device, such as a mobile phone, a computer, a tablet computer, etc., without limitation.

[0200] The application module can be an executable application installed on the electronic device. The application module can call at least one model and provide various interactive functions based on artificial intelligence (AI) technology for the user of the electronic device based on the called model, such as recognizing multimedia data (images, videos, audio) input by the user, generating multimedia data based on the user's input, and conducting human-computer dialogue based on the user's input.

[0201] After receiving the input data, the application module may determine which interactive functions need to be provided to the user based on the input data, and then determine a model in at least one model used to implement these interactive functions as a target model.

[0202] Exemplarily, when it is determined that an image or video needs to be generated, a model with generation capabilities is determined as the target model. When it is determined that the input data is problem data, the application module can determine that a human-computer dialogue function needs to be provided, and then determine a model with dialogue capabilities as the target model, for example, determining the large language model of the aforementioned embodiment as the target model.

[0203] Optionally, when it is determined that the input data is problem data, the application module may also determine different target models according to different problem data. For example, for problem data in different languages, a model suitable for processing problems in the corresponding language may be determined as the target model; and for problem data in different fields, a model suitable for processing problems in the corresponding field may be determined as the target model.

[0204] Taking a target model with conversation capabilities as an example, in this embodiment, when the application module determines that the data input by the user is question data, it can use this question data as the first question data of the current session, and then obtain the conversation data of the current session and the conversation data of multiple rounds of historical sessions from the user's historical data, determine the target round of the target historical session based on these conversation data, and then obtain the target data corresponding to the target round, and input the first question data and the target data into the target model.

[0205] The target model can receive the first question data and target data input by the application module, and process the first question data according to the target data to determine the output data corresponding to the data input by the user, that is, to determine the above-mentioned first reply data. The target model can then return the first reply data to the application module. After receiving the first reply data, the application module can output the first reply data to the user in various forms such as display and voice.

[0206] Alternatively, when the first question data is question data of the first round of the current conversation, the application module may also input only the first question data into the target model, or input the first question data and the dialogue data of the current conversation into the target model.

[0207] The target model can receive the first question data input by the application module, process the first question data to obtain the first reply data, or receive the first question data input by the application module and the conversation data of the current session, process the first question data based on the conversation data of the current session to obtain the first reply data, and then return the first reply data to the application module. After receiving the first reply data, the application module outputs the first reply data to the user.

[0208] When determining the target round of the target historical conversation, the application module can input the first question data, the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations into the target model, so that the target model processes the first question data, the conversation data of the current conversation and the historical conversation to determine the target round of the target historical conversation.

[0209] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0210] For the convenience of description, the above system or device is described by dividing it into various modules or units according to its functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0211] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly contributed to the prior art in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0212] Finally, it should be noted that, in this article, relational terms such as first, second, third and fourth are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0213] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A data processing method, comprising: Get the first question data entered by the user in the current session; Determine a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations, wherein the conversation data of each round includes a question data and an answer data; Based on the target round of the target historical conversation, obtaining target data corresponding to the target round, wherein the target data is data obtained in the process of processing the question data of the target round to obtain the answer data of the target round; The first question data is processed using the target data to determine first answer data corresponding to the first question data.

2. The method according to claim 1, wherein obtaining the conversation data of the current conversation comprises: The conversation data before the first question data in the current conversation, or the conversation data of the historical conversation that meets specific conditions, is obtained as the conversation data of the current conversation, wherein the conversation data of the current conversation includes N rounds of conversations, where N is greater than or equal to 1.

3. The method according to claim 1, wherein determining a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations comprises: Compare the N rounds of conversation data in the current session with multiple rounds of conversation data in the historical session. If there are N rounds of conversation data in the historical session that match the N rounds of conversation data in the current session, determine the historical session as the target historical session, and obtain the target round of the target historical session, where N is greater than or equal to 1.

4. The method according to claim 1, wherein determining a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations comprises: Based on the first question data of the current session, the conversation data of the current session, and the conversation data of multiple rounds of historical sessions, a target round of the target historical session is determined.

5. The method according to claim 4, wherein determining the target round of the target historical conversation based on the first question data of the current conversation, the conversation data of the current conversation, and the conversation data of multiple rounds of historical conversations comprises: In the case where second question data matching the first question data is found in the historical session, a round corresponding to the second question data is used as a target round, and target data of the target round is obtained, wherein the target data of the round corresponding to the second question data is data obtained in the process of obtaining second answer data from the second question data; When no second question data matching the first question data is found in the historical conversation, the conversation data in the current conversation is compared with conversation data of multiple rounds in the historical conversation to determine a target round of the target historical conversation.

6. The method according to claim 1, further comprising: Determine the hash value of each round of dialogue in the session, and / or the hash value of the question data in each round, wherein the hash value is used to determine the comparison between the dialogue data of the current session and the dialogue data of multiple rounds of historical sessions, and / or the comparison between the question data in the current session and the question data of multiple rounds of historical sessions.

7. The method according to claim 1, further comprising: In a case where the first question data is question data of the first round of the current conversation, processing the first question data and determining first answer data corresponding to the first question data; Alternatively, the first question data is processed using the conversation data of the historical conversation that meets specific conditions to determine the first answer data corresponding to the first question data.

8. The method according to claim 1, further comprising: Get the subject of the current session; A historical session whose corresponding topics in the plurality of historical sessions match the topic of the current session is determined as a target historical session.

9. A data processing device, comprising: An input unit, used to obtain first question data input by a user in a current session; a determination unit, configured to determine a target round of a target historical conversation based on the conversation data of the current conversation and the conversation data of multiple rounds of historical conversations, wherein the conversation data of each round includes a question data and an answer data; An acquisition unit, configured to acquire target data corresponding to the target round based on the target round of the target historical conversation, wherein the target data is data obtained in the process of processing the question data of the target round to obtain the answer data of the target round; A processing unit is used to process the first question data using the target data to determine first answer data corresponding to the first question data.

10. An electronic device comprising an application module and at least one model; The application module is used to receive data input by a user, and determine a target model to be called based on analyzing the data input by the user. The application module can call the at least one model; The target model is used to obtain data input by a user, determine output data corresponding to the data input by the user based on the user's historical data and the target data corresponding to the historical data, and return the output data to the application module so that the application module outputs the output data to the user, wherein the historical data includes input data input by the user and output data corresponding to the input data, and the target data corresponding to the historical data is data obtained in the process of processing the input data to obtain the output data.

Images