A dialogue information processing method, device and equipment and storage medium

Abstract

The application provides a dialogue information processing method and device, equipment and a storage medium, wherein the method comprises: obtaining first dialogue information and second dialogue information, the first dialogue information being non-first-round dialogue information in multi-round dialogue information, and the second dialogue information being at least one round of dialogue information before the first dialogue information in the multi-round dialogue information; according to the first dialogue information and the second dialogue information, detecting whether there is information missing in the first dialogue information, if so, completing the first dialogue information by obtaining the missing information in the first dialogue information from the second dialogue information. The application can detect whether there is information missing in the first dialogue information, and when detecting that there is information missing in the first dialogue information, the first dialogue information can be completed, which makes subsequent intent recognition based on the completed dialogue information be able to obtain a more accurate recognition result, and then a correct feedback can be made based on the accurate recognition result, and the user experience is better.

Description

Technical Field

[0001] This application relates to the field of dialogue information processing technology, and in particular to a dialogue information processing method, apparatus, device and storage medium. Background Technology

[0002] With the increasing maturity of the information technology industry and artificial intelligence technology, more and more jobs (such as human customer service and human quality inspection) are being performed by machines. Machines are able to perform these tasks primarily through dialogue systems. These systems analyze dialogue information, detect the user's intent, and then provide feedback. Currently, machines can accurately analyze user intent in single-turn dialogue tasks. However, in multi-turn dialogue tasks, they struggle to accurately analyze user intent, leading to difficulties in providing correct feedback and a poor user experience. Summary of the Invention

[0003] In view of this, this application provides a method, apparatus, device, and storage medium for processing dialogue information, so as to accurately analyze the user's intent based on the processed dialogue information. The technical solution is as follows:

[0004] A dialogue information processing method, comprising:

[0005] Obtain first dialogue information and second dialogue information, wherein the first dialogue information is non-first round dialogue information in multi-round dialogue information, and the second dialogue information is at least one round dialogue information that precedes the first dialogue information in the multi-round dialogue information;

[0006] Based on the first dialogue information and the second dialogue information, detect whether there is any missing information in the first dialogue information;

[0007] When missing information is detected in the first dialogue information, the missing information is obtained from the second dialogue information to complete the first dialogue information.

[0008] Optionally, detecting whether there is missing information in the first dialogue information based on the first dialogue information and the second dialogue information includes:

[0009] Based on the first dialogue information and the second dialogue information, predict the probability that each information unit in the first dialogue information is missing historical context information;

[0010] Based on the predicted probability, determine whether each information unit in the first dialogue information is missing historical context.

[0011] Optionally, the step of detecting whether there is missing information in the first dialogue information based on the first dialogue information and the second dialogue information, and, when missing information is detected in the first dialogue information, obtaining the missing information from the second dialogue information to complete the first dialogue information, includes:

[0012] The first dialogue information and the second dialogue information are processed using a pre-trained dialogue processing model to obtain the completed dialogue information corresponding to the first dialogue information.

[0013] The dialogue processing model has the ability to detect and complete missing information in the first dialogue information.

[0014] Optionally, the step of processing the first dialogue information and the second dialogue information using a pre-trained dialogue processing model to obtain the completed dialogue information corresponding to the first dialogue information includes:

[0015] The dialogue processing model is used to process the first dialogue information and the second dialogue information to determine and output the information units to be output one by one.

[0016] The process of determining a unit of information to be output using the dialogue processing model includes:

[0017] The probability that the information unit to be output is missing historical context is predicted using the dialogue processing model, and the information unit to be output is determined based on the predicted probability.

[0018] Optionally, determining the information unit to be output based on the predicted probability includes:

[0019] If the predicted probability is greater than or equal to the preset probability threshold, then select an information unit from the second dialogue information as the information unit to be output.

[0020] If the predicted probability is less than the probability threshold, then an information unit is obtained from the first dialogue information and used as the information unit to be output.

[0021] Optionally, determining the information unit to be output based on the probability that the information unit to be output is missing historical context includes:

[0022] If the predicted probability is greater than or equal to the preset first probability threshold, then select an information unit from the second dialogue information as the information unit to be output.

[0023] If the predicted probability is less than or equal to the preset second probability threshold, then the information unit is obtained from the first dialogue information and used as the information unit to be output.

[0024] If the predicted probability is greater than the second probability threshold and less than the first probability threshold, then an information unit is generated based on the output information unit and the dictionary, and used as the information unit to be output.

[0025] Optionally, the dialogue processing model includes an encoding layer and a decoding layer;

[0026] The encoding layer encodes the embedded representation of the first dialogue information and the embedded representation of the second dialogue information to obtain the encoding results corresponding to the first dialogue information and the second dialogue information respectively;

[0027] At each decoding moment, the decoding layer predicts the probability that the information unit to be output at the current decoding moment is missing historical context information, based on the encoding results corresponding to the first dialogue information and the second dialogue information, respectively, and the information unit prediction vector of the previous decoding moment. The layer then determines the information unit to be output based on the predicted probability and outputs it.

[0028] Optionally, the process of obtaining the embedded representation of dialogue information includes:

[0029] Obtain the representation vector of each information unit in the dialogue information, the position representation vector of each information unit in the dialogue information, and the dialogue round representation vector of the dialogue information;

[0030] The representation vector of each information unit in the dialogue information, the position representation vector of each information unit in the dialogue information, and the dialogue turn number representation vector of the dialogue information are fused together, and the fused vector is used as the embedded representation of the dialogue information.

[0031] Optionally, the decoding layer includes: a first attention module, a second attention module, and a missing information detection and completion module;

[0032] The first attention module determines the information unit prediction vector for the current decoding time based on the information unit prediction vector from the previous decoding time.

[0033] The second attention module determines a first vector that can characterize the correlation between the information unit to be output at the current decoding time and each information unit in the first dialogue information, and a second vector that can characterize the correlation between the information unit to be output at the current decoding time and each information unit in the second dialogue information, based on the encoding results corresponding to the first dialogue information and the information unit prediction vector at the current decoding time.

[0034] The missing information detection and completion module predicts the probability that the information unit to be output at the current decoding moment is missing historical information based on the first vector and the second vector, and outputs the information unit according to the predicted probability.

[0035] Optionally, the second attention module includes: an attention module corresponding to the first dialogue information and an attention module corresponding to the second dialogue information;

[0036] The attention module corresponding to the first dialogue information determines the weight of each information unit in the first dialogue information based on the information unit prediction vector at the current decoding time and the encoding result corresponding to the first dialogue information, and determines the first vector based on the determined weight and the encoding result corresponding to the first dialogue information.

[0037] The attention module corresponding to the second dialogue information determines the weight of each information unit in the second dialogue information based on the information unit prediction vector at the current decoding time and the encoding result corresponding to the second dialogue information, and determines the second vector based on the determined weight and the encoding result corresponding to the second dialogue information.

[0038] Optionally, the training process of the dialogue processing model includes:

[0039] Training data is obtained from the training dataset, wherein the training data includes a first dialogue sample and a second dialogue sample, and the first dialogue sample is labeled with dialogue information without missing information corresponding to the first dialogue sample.

[0040] The first dialogue sample and the second dialogue information are processed using a dialogue processing model to predict the probability distribution corresponding to each information unit to be output. When predicting the probability distribution corresponding to an information unit to be output, the probability that the information unit to be output is missing historical information is predicted as the missing probability corresponding to the information unit to be output. Based on the missing probability corresponding to the information unit to be output, each candidate information unit corresponding to the information unit to be output is determined, and the probability that the information unit to be output is each of the corresponding candidate information units is predicted to obtain the probability distribution corresponding to the information unit to be output.

[0041] Based on the probability distribution corresponding to each information unit to be output and the annotation information of the first dialogue sample, the parameters of the dialogue processing model are updated.

[0042] Optionally, determining each candidate information unit corresponding to the information unit to be output based on the missing probability of the information unit to be output includes:

[0043] If the missing probability corresponding to the information unit to be output is greater than or equal to the preset first probability threshold, then each information unit in the second dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output.

[0044] If the missing probability corresponding to the information unit to be output is less than or equal to the preset second probability threshold, then each information unit in the first dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output.

[0045] If the missing probability corresponding to the information unit to be output is greater than the second probability threshold and less than the first probability threshold, then each word in the dictionary is determined to be a candidate information unit corresponding to the information unit to be output.

[0046] A dialogue information processing device includes: a dialogue information acquisition module, an information missing detection module, and a dialogue information completion module;

[0047] The dialogue information acquisition module is used to acquire first dialogue information and second dialogue information, wherein the first dialogue information is non-first round dialogue information in multi-round dialogue information, and the second dialogue information is at least one round of dialogue information that precedes the first dialogue information in the multi-round dialogue information.

[0048] The information missing detection module is used to detect whether there is any missing information in the first dialogue information based on the first dialogue information and the second dialogue information.

[0049] The dialogue information completion module is used to obtain the missing information in the first dialogue information from the second dialogue information and complete the first dialogue information when the information missing detection module detects that there is missing information in the first dialogue information.

[0050] A dialogue information processing device includes: a memory and a processor;

[0051] The memory is used to store programs;

[0052] The processor is configured to execute the program to implement each step of the dialogue information processing method described above.

[0053] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the various steps of the dialogue information processing method described in any of the preceding claims.

[0054] As can be seen from the above solution, the dialogue information processing method, apparatus, device, and storage medium provided in this application first acquire first dialogue information (non-first-round dialogue information in multi-round dialogue information) and second dialogue information (at least one round of dialogue information preceding the first dialogue information in multi-round dialogue information). Then, based on the first and second dialogue information, it detects whether there is missing information in the first dialogue information. When missing information is detected, the missing information in the first dialogue information is obtained from the second dialogue information to complete the first dialogue information. The dialogue information processing method provided in this application can detect whether there is missing information in the first dialogue information, and when missing information is detected, it can obtain the missing information in the first dialogue information from the second dialogue information to complete the first dialogue information. This allows for more accurate recognition results when subsequent intent recognition is performed based on the completed dialogue information, thereby enabling correct feedback based on accurate recognition results and a better user experience. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0056] Figure 1 A flowchart illustrating the dialogue information processing method provided in an embodiment of this application;

[0057] Figure 2 A schematic diagram of the structure of the dialogue processing model provided in the embodiments of this application;

[0058] Figure 3 A schematic diagram of the training process of the dialogue processing model provided in the embodiments of this application;

[0059] Figure 4 This is a schematic diagram of the structure of the dialogue information processing device provided in the embodiments of this application;

[0060] Figure 5 This is a schematic diagram of the structure of the dialogue information processing device provided in the embodiments of this application. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] In the process of implementing this application, the applicant discovered that: in single-turn dialogue tasks, the dialogue information is relatively complete, and the dialogue system can accurately detect the user's intent based on the dialogue information. However, in multi-turn dialogue tasks, non-first-turn dialogue information (such as the current-turn dialogue information) often omits some information mentioned in previous-turn dialogue information. Since most current dialogue systems perform intent analysis based on one-turn dialogue information (such as the current-turn dialogue information), the omission of information in the current-turn dialogue information will cause a significant deviation between the intent obtained by the dialogue system based on the current dialogue information and the true intent, thereby reducing the user experience.

[0063] In view of this, the applicant sought to propose a scheme capable of detecting and completing missing information in dialogue information, so as to accurately analyze intent subsequently. To this end, the applicant conducted research and ultimately proposed a dialogue information processing method with good performance. Under multi-turn dialogue tasks, this scheme can detect and complete missing information in dialogue information. The dialogue information processing method provided in this application can be applied to electronic devices with data processing capabilities. This electronic device can be a network-side server (which can be a single server, multiple servers, or a server cluster) or a user-side terminal, such as a PC, laptop, PAD, smartphone, smart home appliance, learning machine, smart vehicle terminal, etc. The dialogue information processing method provided in this application will be described below through the following embodiments.

[0064] First Embodiment

[0065] Please see Figure 1 The diagram illustrates a flowchart of a dialogue information processing method provided in an embodiment of this application. The method may include:

[0066] Step S101: Obtain the first dialogue information and the second dialogue information.

[0067] Wherein, the first dialogue information is non-first dialogue information in the multi-turn dialogue information, and the second dialogue information is at least one round of dialogue information that precedes the first dialogue information in the multi-turn dialogue information. Preferably, the second dialogue information is at least one round of dialogue information that precedes the first dialogue information and is closest to the first dialogue information in the multi-turn dialogue information.

[0068] In some real-time application scenarios, the first dialogue information can be the current round of dialogue information, that is, the latest round of dialogue information in a multi-round dialogue. The second dialogue information is preferably the round of dialogue information most recent to the first dialogue information or the multi-round dialogue information most recent to the first dialogue information. For example, if the multi-round dialogue information consists of 4 rounds of dialogue information, and the 4th round of dialogue information is the current round of dialogue information, then the first dialogue information is the 4th round of dialogue information, and the second dialogue information can be the 3rd round of dialogue information, or the 2nd to 3rd rounds of dialogue information, or the 1st to 3rd rounds of dialogue information.

[0069] In some non-real-time application scenarios, the first dialogue information can be a non-first round of dialogue information in a multi-turn dialogue, and the second dialogue information is preferably the most recent round of dialogue information or a multi-turn dialogue information in the multi-turn dialogue information. For example, if the multi-turn dialogue information consists of 4 rounds of dialogue, then the first dialogue information can be the 4th round of dialogue, or the 3rd round of dialogue, or the 2nd round of dialogue. If the first dialogue information is the 4th round of dialogue, then the second dialogue information can be the 3rd round of dialogue, or the 2nd to 3rd rounds of dialogue, or the 1st to 3rd rounds of dialogue. If the first dialogue information is the 3rd round of dialogue, then the second dialogue information can be the 2nd round of dialogue, or the 1st to 2nd rounds of dialogue. If the first dialogue information is the 2nd round of dialogue, then the second dialogue information is the 1st round of dialogue.

[0070] Step S102: Based on the first dialogue information and the second dialogue information, detect whether there is any missing information in the first dialogue information.

[0071] Specifically, based on the first dialogue information and the second dialogue information, the probability that each information unit in the first dialogue information lacks historical context can be predicted, and then the lack of historical context can be determined based on the predicted probability. More specifically, for any information unit in the first dialogue information, if the predicted probability that the information unit lacks historical context is greater than or equal to a preset probability threshold, then it can be determined that the information unit lacks historical context.

[0072] Among them, the information unit can be a word-level information unit, that is, the first dialogue information can be segmented into words to obtain each word contained in the first dialogue information, and each word is an information unit. The information unit can also be a character-level information unit, that is, the first dialogue information can be segmented into characters, and each segmented character is an information unit.

[0073] Step S103: When missing information is detected, retrieve the missing information from the first dialogue information from the second dialogue information and complete the first dialogue information.

[0074] Specifically, for each information unit in the first dialogue information, if it is determined that the information unit is missing historical information, the missing information is obtained from the second dialogue information, and the information unit is completed using the obtained information.

[0075] For example, multi-turn dialogue information is as follows:

[0076] "Role 1: Which movie do you like the most?"

[0077] Character 2: Farewell My Concubine

[0078] Character 1: What a coincidence, I like it too!

[0079] In the example above, the first dialogue information could be "What a coincidence, I like it too!", and the second dialogue information could be "What's your favorite movie? Farewell My Concubine." For the information unit "like" in the first dialogue information, if the probability that this information unit lacks historical context is greater than the probability threshold, then it can be determined that the information unit "like" lacks historical context. In this case, the missing information for the information unit "like" is obtained from the second dialogue information. Through analysis, it is found that the missing historical context for the information unit "like" is "Farewell My Concubine." Therefore, "Farewell My Concubine" is used to complete the information unit "like," that is, "Farewell My Concubine" is added after the information unit "like." The completed dialogue information is "What a coincidence, I like Farewell My Concubine too!".

[0080] In one possible implementation, steps S102 and S103 described above can be implemented using a pre-trained dialogue processing model. That is, the pre-trained dialogue processing model is used to process the first dialogue information and the second dialogue information to obtain the completed dialogue information corresponding to the first dialogue information. The dialogue processing model has the ability to detect and complete missing information in the first dialogue information.

[0081] The dialogue processing model in this embodiment is trained using training data from the training dataset. The training data includes a first dialogue sample and a second dialogue sample. The first dialogue sample is non-first-round dialogue information in a multi-round dialogue sample. The second dialogue sample is at least one round of dialogue information in a multi-round dialogue sample that precedes the first dialogue sample, preferably at least one round of dialogue information closest to the first dialogue sample (preferably multi-round dialogue information). The first dialogue sample is labeled with dialogue information corresponding to the first dialogue sample that has no missing information.

[0082] In one possible implementation, the first dialogue sample in each training data set can be a dialogue with missing information. In another possible implementation, the first dialogue samples in a portion of the training data can be dialogues with missing information, while the first dialogue samples in another portion of the training data can be dialogues without missing information. It should be noted that if the first dialogue sample is dialogue with missing information, then the annotation information of the first dialogue sample is the dialogue information after the first dialogue sample is completed (i.e., the "dialogue information without missing information corresponding to the first dialogue sample" mentioned above refers to the dialogue information after the first dialogue sample is completed). If the first dialogue sample is dialogue information without missing information, then the annotation information of the first dialogue sample is the first dialogue sample itself (i.e., the "dialogue information without missing information corresponding to the first dialogue sample" mentioned above refers to the first dialogue sample itself).

[0083] The dialogue information processing method provided in this application first acquires first dialogue information (e.g., current round dialogue information) and second dialogue information (e.g., historical round dialogue information). Then, based on the first and second dialogue information, it detects whether there is missing information in the first dialogue information. When missing information is detected, the missing information is retrieved from the second dialogue information to complete the first dialogue information. This dialogue information processing method can detect whether there is missing information in the first dialogue information and, when such missing information is detected, can complete the first dialogue information. This allows for more accurate intent recognition results based on the completed dialogue information, leading to correct feedback and a better user experience.

[0084] Second Embodiment

[0085] The above embodiments mention that a pre-trained dialogue processing model can be used to process the first dialogue information and the second dialogue information to obtain the completed dialogue information corresponding to the first dialogue information. This embodiment describes this process.

[0086] The process of processing first and second dialogue information using a pre-trained dialogue processing model to obtain the completed dialogue information corresponding to the first dialogue information may include: processing the first and second dialogue information using the dialogue processing model to determine and output information units to be output one by one. Specifically, when determining an information unit to be output using the dialogue processing model, the model first predicts the probability that the information unit to be output is missing historical context, and then determines the information unit to be output based on the predicted probability.

[0087] There are several ways to determine the output information unit based on the predicted probability:

[0088] In one possible implementation, a probability threshold P can be preset. th If the predicted probability is greater than or equal to the probability threshold P th If the information unit to be output is determined to be missing historical information, then information units are selected from the second dialogue information. These selected information units are then used as the information units to be output. If the predicted probability is less than the probability threshold P... th If the information unit is obtained from the first dialogue information, it will be used as the information unit to be output. This implementation can extract the information omitted in the first dialogue from the second dialogue information, and retain the information from the first dialogue for the parts that are not omitted.

[0089] The process of selecting information units from the second dialogue information may include: predicting the probability that the information unit to be output is one of the information units in the second dialogue information, so as to obtain the probability corresponding to each information unit in the second dialogue information, and determining the information unit corresponding to the highest probability among the obtained probabilities as the information unit to be output. Similarly, the process of obtaining information units from the first dialogue information may include: predicting the probability that the information unit to be output is one of the information units in the first dialogue information, so as to obtain the probability corresponding to each information unit in the first dialogue information, and determining the information unit corresponding to the highest probability among the obtained probabilities as the information unit to be output.

[0090] In another possible implementation, two probability thresholds P can be preset. th1 and P th2 (P th1 Greater than P th2 If the predicted probability is greater than or equal to the probability threshold P th1 If the information unit to be output is determined to be missing historical information, then information units are selected from the second dialogue information. The information units selected from the second dialogue information are used as the information units to be output. If the predicted probability is less than or equal to the probability threshold P... th2 If the information unit to be output is determined to be not missing historical information, then the information unit is obtained from the first dialogue information and used as the information unit to be output. If the predicted probability is greater than the probability threshold P... th2 And less than the probability threshold P th1Then, information units are generated based on the output information units and the dictionary, and the generated information units serve as the information units to be output. This implementation method can extract information omitted in the first dialogue from the second dialogue information. For the parts that are not omitted, the information from the first dialogue can be retained. It can also achieve a certain degree of dialogue rewriting, making the dialogue more in line with human speaking habits.

[0091] The process of generating information units based on the output information units and the dictionary includes: referring to the output information units, predicting the probability that the information unit to be output is a word in the dictionary, so as to obtain the probability corresponding to each word in the dictionary, and determining the word with the highest probability among the obtained probabilities as the information unit to be output. It should be noted that in the second implementation, the implementation methods of selecting information units from the second dialogue information and obtaining information units from the first dialogue information are the same as those in the first implementation, and will not be described again in this embodiment.

[0092] Optionally, the dialogue processing model in this embodiment can be an Encoder-Decode framework model; please refer to [link / reference]. Figure 2 This diagram illustrates the structure of a dialogue processing model within the Encoder-Decode framework, which includes an input layer, an encoding layer, and a decoding layer. Wherein:

[0093] The input layer 201 of the dialogue processing model is used to input the embedded representation of the first dialogue information and the embedded representation of the second dialogue information.

[0094] The process of obtaining the embedded representation of a dialogue information (assuming it is represented as z) may include: obtaining the representation vector of each information unit in the dialogue information z, the position representation vector of each information unit in the dialogue information z, and the dialogue turn number representation vector of the dialogue information z; fusing the representation vector of each information unit in the dialogue information z, the position representation vector of each information unit in the dialogue information z, and the dialogue turn number representation vector of the dialogue information z, and using the fused vector as the embedded representation of the dialogue information z.

[0095] It should be noted that if the dialogue information z is a one-turn dialogue, then the representation vector, position representation vector, and dialogue turn number representation vector of each information unit in that turn are obtained. For each information unit in that turn, the representation vector, position representation vector, and dialogue turn number representation vector of that information unit are summed to obtain the summed vector corresponding to each information unit in that turn. The summed vectors corresponding to each information unit in that turn form the embedded representation of the dialogue information z. If the dialogue information z is a multi-turn dialogue, then the above method (the method for obtaining the embedded representation of a one-turn dialogue) is used to obtain the embedded representation of each turn of the dialogue. The embedded representations of each turn of the dialogue are concatenated to obtain the embedded representation of the dialogue information z. It should be noted that the dialogue turn number of a one-turn dialogue refers to which turn of the dialogue in a multi-turn dialogue.

[0096] The encoding layer 202 of the dialogue processing model encodes the embedded representations of the first dialogue information and the second dialogue information to obtain the encoding results corresponding to the first and second dialogue information, respectively. Assume the first dialogue information is represented by U... n Let H represent the second dialogue information. Then, the encoding result corresponding to the first dialogue information can be represented as Em(U). n The encoding result corresponding to the second dialogue information can be represented as Em(H).

[0097] In this model, the encoding layer 202 performs feature transformation on the embedded representation of the first dialogue information. The transformed features include the historical and future information of each information unit in the first dialogue information, i.e., the information of the information units preceding and following each information unit. Similarly, the encoding layer 202 performs feature transformation on the embedded representation of the second dialogue information. The transformed features include the historical and future information of each information unit in the second dialogue information. That is, the encoding result Em(U) corresponding to the first dialogue information... n ) is the context vector corresponding to the first dialogue information, and Em(H) is the encoding result corresponding to the second dialogue information.

[0098] Optionally, the encoding layer 202 of the dialogue processing model may, but is not limited to, using a BERT language model, an RNN model, an LSTM model, etc. It should be noted that if the encoding layer 202 uses a BERT language model, the second dialogue information needs to be limited to 512 characters. If the encoding layer 202 uses an RNN model or an LSTM model, there is no limit to the total number of characters in the second dialogue information. Preferably, even if the encoding layer does not limit the total number of characters in the second dialogue information, in order to improve processing efficiency, the second dialogue information can be limited to a certain number of characters (such as 512 characters).

[0099] At each decoding time, the decoding layer 203 of the dialogue processing model predicts the probability that the information unit to be output at the current decoding time is missing historical context information based on the encoding results corresponding to the first dialogue information and the second dialogue information, as well as the information unit prediction vector of the previous decoding time. The output information unit is then determined and output based on the predicted probability.

[0100] Specifically, at each decoding time, the decoding layer 203 of the dialogue processing model first determines the information unit prediction vector for the current decoding time based on the information unit prediction vector from the previous decoding time. Then, it determines the information unit prediction vector based on the encoding results corresponding to the first and second dialogue information (i.e., Em(U...). n Based on the information unit prediction vector N at the current decoding time, and Em(H)), a first vector Y(U) is determined that can characterize the correlation between the information unit to be output at the current decoding time and each information unit in the first dialogue information. n ), and a second vector Y(H) that can characterize the correlation between the information unit to be output at the current decoding moment and each information unit in the second dialogue information, and finally based on the first vector Y(U) n The second vector Y(H) predicts the probability λ that the information unit to be output at the current decoding time is missing historical context, and outputs the information unit according to the predicted probability λ.

[0101] In one possible implementation, the decoding layer of the dialogue processing model may include: a first attention module 2031, a second attention module 2032, and a missing information detection and completion module 2033.

[0102] At each decoding moment, the first attention module 2031 determines the information unit prediction vector N for the current decoding moment based on the information unit prediction vector X from the previous decoding moment, and the second attention module 2032 determines the information unit prediction vector N based on the encoding result Em(U) corresponding to the first dialogue information. n The first vector Y(U) is determined by the information unit prediction vector N at the current decoding time. n), and, based on the encoding result Em(H) corresponding to the second dialogue information and the information unit prediction vector N at the current decoding time, the above-mentioned second vector Y(H) is determined, and the missing information detection and completion module block 2033 determines the above-mentioned second vector Y(H) based on the first vector Y(H). n The second vector Y(H) predicts the probability λ that the information unit to be output at the current decoding time is missing historical context, and determines and outputs the information unit to be output based on the predicted probability λ. The implementation process of determining the information unit to be output based on the predicted probability λ is described in the two implementation methods given above.

[0103] There are multiple ways to implement the second attention module 2032. In one possible implementation, the second attention module 2032 may include an attention module 2032-1 corresponding to the first dialogue information and an attention module 2032-2 corresponding to the second dialogue information.

[0104] The attention module 2032-1 corresponding to the first dialogue information predicts the information unit prediction vector N and the encoding result Em(U) corresponding to the first dialogue information based on the information unit prediction vector N at the current decoding time. n ), determine the weight corresponding to each information unit in the first dialogue information, and based on the determined weight and the encoding result Em(U) corresponding to the first dialogue information. n Determine the first vector Y(U) n (For example, the vectors corresponding to each information unit in the encoding result corresponding to the first dialogue information can be weighted and summed according to the determined weights); Similarly, the attention module 2032-2 corresponding to the second dialogue information determines the weights corresponding to each information unit in the second dialogue information based on the information unit prediction vector N at the current decoding time and the encoding result Em(H) corresponding to the second dialogue information, and determines the second vector Y(H) based on the determined weights and the encoding result Em(H) corresponding to the second dialogue information (for example, the vectors corresponding to each information unit in the encoding result corresponding to the second dialogue information can be weighted and summed according to the determined weights).

[0105] Optionally, the attention module 2032-1 corresponding to the first dialogue information predicts the vector N of the information unit at the current decoding time and the encoding result Em(U) corresponding to the first dialogue information. n The process of determining the weight corresponding to each information unit in the first dialogue information may include: calculating the information unit prediction vector N at the current decoding time and the encoding result Em(U) corresponding to the first dialogue information. nThe similarity (e.g., cosine similarity) of the vectors corresponding to each information unit in the first dialogue information is used as the weight of each information unit in the first dialogue information. The process of determining the weight of each information unit in the second dialogue information by the attention module 2032-2 based on the information unit prediction vector N at the current decoding time and the encoding result Em(H) corresponding to the second dialogue information is similar and will not be described in detail here.

[0106] In another possible implementation, the second attention module may include a shared attention module, namely the first vector Y(U) mentioned above. n The determination of the first vector Y(U) and the second vector Y(H) are both determined by this shared attention module, i.e., the determination of the first vector Y(U) and the second vector Y(H) are both determined by this shared attention module. n The determination of the first vector Y(H) and the second vector Y(H) share an attention module.

[0107] After obtaining the first vector Y(U) n After obtaining the first vector Y(U) and the second vector Y(H), the missing information detection and completion module 2033, based on the first vector Y(U), performs the missing information detection and completion based on the second vector Y(H). n The first vector Y(U) and the second vector Y(H) predict the probability λ that the output information unit at the current decoding time is missing historical context. Specifically, the first vector Y(U) is first... n The first vector Y(U) is concatenated with the second vector Y(H), and then the concatenated vector is fed into the feedforward neural network. Next, the first vector Y(U) is processed... n The second vector Y(H) and the output D of the feedforward neural network are linearly transformed, and the probability λ of the information unit to be output at the current decoding time being the missing historical information is determined based on the linear transformation result.

[0108] The output D of the feedforward neural network can be expressed as:

[0109] D=FNN([Y(H)||Y(Un)]) (1)

[0110] Where Y(H)||Y(Un) represents the first vector Y(U) n The concatenated vector of the first vector and the second vector Y(H), FNN represents a feedforward neural network.

[0111] The probability λ that the output information unit at the current decoding time t is missing historical context can be calculated using the following formula:

[0112] λ=σ(ω d D t +ω d Y(H) t +ω H Y(U n ) t (2)

[0113] Where σ is the sigmoid function, ω d and ω H The parameters ω are all obtained through learning. d D t +ω d Y(H) t +ω H Y(U n ) t For the first vector Y(U) n The transformation result is obtained by linearly transforming the second vector Y(H) and the output D of the feedforward neural network.

[0114] After obtaining the probability λ that the information unit to be output at the current decoding moment is missing historical context, the missing information detection and completion module determines and outputs the information unit to be output at the current decoding moment according to one of the two implementation methods provided above, based on the obtained probability λ. That is, if λ≥P th If the information unit to be output at the current decoding moment is determined to be missing historical context information, then information units are selected from the second dialogue information and output. If λ <P th If λ ≥ P, then the information unit is obtained from the first dialogue information and output; or, if λ ≥ P th1 If the information unit to be output at the current decoding moment is determined to be the missing historical context, the information unit is selected from the second dialogue information and output, if λ≤P th2 If the information unit to be output at the current decoding moment is determined to be not missing historical information, then the information unit is obtained from the first dialogue information and output. If P th2 <λ <P th1 Then, based on the already output information units and the dictionary, information units are generated and output.

[0115] Third Embodiment

[0116] This embodiment describes the training process of the dialogue processing model.

[0117] Please see Figure 3 The diagram illustrates the process of training a dialogue processing model, which may include:

[0118] Step S301: Obtain training data S from the training dataset.

[0119] As mentioned in the above embodiments, the training data S includes a first dialogue sample s1 and a second dialogue sample s2, wherein the first dialogue sample s1 is labeled with dialogue information without missing information corresponding to the first dialogue sample s1.

[0120] Step S302: Use the dialogue processing model to process the first dialogue sample s1 and the second dialogue sample s2 to predict the probability distribution corresponding to each information unit to be output.

[0121] The process of predicting the probability distribution corresponding to each information unit to be output includes:

[0122] Step a1: Predict the probability that the information unit to be output is missing historical information, and use it as the missing probability corresponding to the information unit to be output.

[0123] Specifically, the encoding layer of the dialogue processing model is used to encode the embedding representation of the first dialogue sample s1 and the embedding representation of the second dialogue sample s2 to obtain the encoding results corresponding to the first dialogue sample s1 and the second dialogue sample s2 respectively; at each decoding time, the decoding layer of the dialogue processing model is used to predict the probability that the information unit to be output at the current decoding time is missing historical context information, based on the encoding results corresponding to the first dialogue sample s1 and the second dialogue sample s2 respectively and the information unit prediction vector of the previous decoding time.

[0124] In this embodiment, the specific implementation process of the decoding layer of the dialogue processing model predicting the probability that the output information unit at the current decoding time is missing historical context information at each decoding time, based on the encoding results corresponding to the first dialogue sample s1 and the second dialogue sample s2 respectively and the information unit prediction vector of the previous decoding time, is similar to the implementation process of the decoding layer of the dialogue processing model at each decoding time, based on the encoding results corresponding to the first dialogue information and the second dialogue information respectively and the information unit prediction vector of the previous decoding time. For details, please refer to the above embodiments, and this embodiment will not be repeated here.

[0125] Step a2: Based on the missing probability corresponding to the information unit to be output, determine each candidate information unit corresponding to the information unit to be output, and predict the probability that the information unit to be output is one of the corresponding candidate information units, so as to obtain the probability distribution corresponding to the information unit to be output.

[0126] There are several ways to implement step a2:

[0127] In one possible implementation, if the missing probability corresponding to the information unit to be output is greater than or equal to a preset probability threshold P... th Then, each information unit in the second dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output. The probability that the information unit to be output is one of the information units in the second dialogue sample is then determined to obtain the probability distribution corresponding to the information unit to be output. If the missing probability corresponding to the information unit to be output is less than the probability threshold P...th Then, each information unit in the first dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output, and the probability of the information unit to be output being each information unit in the first dialogue sample is determined to obtain the probability distribution corresponding to the information unit to be output.

[0128] In one possible implementation, if the missing probability corresponding to the information unit to be output is greater than or equal to a preset probability threshold P... th1 Then, each information unit in the second dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output. The probability that the information unit to be output is one of the information units in the second dialogue sample is then determined to obtain the probability distribution corresponding to the information unit to be output. If the missing probability corresponding to the information unit to be output is less than or equal to the probability threshold P... th2 Then, each information unit in the first dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output. The probability that the information unit to be output is one of the information units in the first dialogue sample is then determined to obtain the probability distribution corresponding to the information unit to be output. If the missing probability corresponding to the information unit to be output is greater than a preset second probability threshold P... th2 And less than the preset probability threshold P th1 Then, each word in the dictionary is determined to be a candidate information unit corresponding to the information unit to be output, and the probability of the information unit to be output being a word in the dictionary is determined to obtain the probability distribution corresponding to the information unit to be output.

[0129] It should be noted that if the probability distribution corresponding to the information unit to be output is determined using the first implementation method described above during the training phase, then in the second embodiment described above, after obtaining λ, the probability distribution is determined by comparing λ with P. th A comparison is made to further determine the information unit to be output. If the probability distribution corresponding to the information unit to be output is determined using the second implementation method described above during the training phase, then in the second embodiment described above, after obtaining λ, λ is compared with P... th1 and P th2 The comparisons are made to further determine the information units to be output.

[0130] Step S303: Determine the prediction loss of the dialogue processing model based on the probability distribution corresponding to each information unit to be output and the annotation information of the first dialogue sample.

[0131] Optionally, the cross-entropy loss can be determined based on the probability distribution corresponding to each information unit to be output and the annotation information of the first dialogue sample, and used as the prediction loss of the dialogue processing model.

[0132] Step S304: Update the parameters of the dialogue processing model based on the prediction loss of the dialogue processing model.

[0133] The dialogue processing model is trained iteratively multiple times according to the above process until the training termination condition is met. The model obtained after training is the dialogue processing model that can detect and complete missing dialogue information.

[0134] Fourth embodiment

[0135] This application also provides a dialogue information processing device. The dialogue information processing device provided in this application is described below. The dialogue information processing device described below can be referred to in correspondence with the dialogue information processing method described above.

[0136] Please see Figure 4 The diagram shows a schematic of the structure of the dialogue information processing device provided in the embodiment of this application, which may include: a dialogue information acquisition module 401, an information missing detection module 402, and a dialogue information completion module 403.

[0137] The dialogue information acquisition module 401 is used to acquire the first dialogue information and the second dialogue information.

[0138] Wherein, the first dialogue information is non-first-round dialogue information in the multi-round dialogue information, and the second dialogue information is at least one round of dialogue information that precedes the first dialogue information in the multi-round dialogue information.

[0139] The information missing detection module 402 is used to detect whether there is information missing in the first dialogue information based on the first dialogue information and the second dialogue information;

[0140] The dialogue information completion module 403 is used to complete the first dialogue information by obtaining the missing information from the second dialogue information when the information missing detection module detects that there is missing information in the first dialogue information.

[0141] Optionally, the information missing detection module 402 is specifically used to predict the probability that each information unit in the first dialogue information is missing historical context information based on the first dialogue information and the second dialogue information, and to determine whether each information unit in the first dialogue information is missing historical context information based on the predicted probability.

[0142] Optionally, the information missing detection module 402 and the dialogue information completion module 403 are implemented by a dialogue processing model. The dialogue processing model has the capability to detect and complete missing information in the first dialogue information.

[0143] Optionally, the dialogue processing model processes the first dialogue information and the second dialogue information to determine and output the information units to be output one by one;

[0144] The process of the dialogue processing model determining an output information unit includes: the dialogue processing model predicts the probability that the output information unit is missing historical context, and determines the output information unit based on the predicted probability.

[0145] Optionally, when the dialogue processing model determines the information unit to be output based on the predicted probability, it is specifically used for:

[0146] If the predicted probability is greater than or equal to the preset probability threshold, then select an information unit from the second dialogue information as the information unit to be output.

[0147] If the predicted probability is less than the probability threshold, then an information unit is obtained from the first dialogue information and used as the information unit to be output.

[0148] Optionally, when the dialogue processing model determines the information unit to be output based on the predicted probability, it is specifically used for:

[0149] If the predicted probability is greater than or equal to the preset first probability threshold, then select an information unit from the second dialogue information as the information unit to be output.

[0150] If the predicted probability is less than or equal to the preset second probability threshold, then the information unit is obtained from the first dialogue information and used as the information unit to be output.

[0151] If the predicted probability is greater than the second probability threshold and less than the first probability threshold, then an information unit is generated based on the output information unit and the dictionary, and used as the information unit to be output.

[0152] Optionally, the dialogue processing model includes an encoding layer and a decoding layer;

[0153] The encoding layer encodes the embedded representation of the first dialogue information and the embedded representation of the second dialogue information to obtain the encoding results corresponding to the first dialogue information and the second dialogue information respectively;

[0154] At each decoding moment, the decoding layer predicts the probability that the information unit to be output at the current decoding moment is missing historical context information, based on the encoding results corresponding to the first dialogue information and the second dialogue information, respectively, and the information unit prediction vector of the previous decoding moment. The layer then determines the information unit to be output based on the predicted probability and outputs it.

[0155] The dialogue information processing apparatus provided in this application embodiment may further include: a dialogue embedding representation acquisition module. When acquiring the embedding representation of dialogue information, the dialogue embedding representation acquisition module is used to:

[0156] Obtain the representation vector of each information unit in the dialogue information, the position representation vector of each information unit in the dialogue information, and the dialogue round representation vector of the dialogue information;

[0157] The representation vector of each information unit in the dialogue information, the position representation vector of each information unit in the dialogue information, and the dialogue turn number representation vector of the dialogue information are fused together, and the fused vector is used as the embedded representation of the dialogue information.

[0158] Optionally, the decoding layer of the dialogue processing model may include: a first attention module, a second attention module, and a missing information detection and completion module;

[0159] The first attention module determines the information unit prediction vector for the current decoding time based on the information unit prediction vector from the previous decoding time.

[0160] The second attention module determines a first vector that can characterize the correlation between the information unit to be output at the current decoding time and each information unit in the first dialogue information, and a second vector that can characterize the correlation between the information unit to be output at the current decoding time and each information unit in the second dialogue information, based on the encoding results corresponding to the first dialogue information and the information unit prediction vector at the current decoding time.

[0161] The missing information detection and completion module predicts the probability that the information unit to be output at the current decoding moment is missing historical information based on the first vector and the second vector, and outputs the information unit according to the predicted probability.

[0162] Optionally, the second attention module includes: an attention module corresponding to the first dialogue information and an attention module corresponding to the second dialogue information;

[0163] The attention module corresponding to the first dialogue information determines the weight of each information unit in the first dialogue information based on the information unit prediction vector at the current decoding time and the encoding result corresponding to the first dialogue information, and determines the first vector based on the determined weight and the encoding result corresponding to the first dialogue information.

[0164] The attention module corresponding to the second dialogue information determines the weight of each information unit in the second dialogue information based on the information unit prediction vector at the current decoding time and the encoding result corresponding to the second dialogue information, and determines the second vector based on the determined weight and the encoding result corresponding to the second dialogue information.

[0165] The dialogue information processing device provided in this application embodiment may further include: a model training module.

[0166] The model training module is used for:

[0167] Training data is obtained from the training dataset, wherein the training data includes a first dialogue sample and a second dialogue sample, and the first dialogue sample is labeled with dialogue information without missing information corresponding to the first dialogue sample.

[0168] The first dialogue sample and the second dialogue information are processed using a dialogue processing model to predict the probability distribution corresponding to each information unit to be output. When predicting the probability distribution corresponding to an information unit to be output, the probability that the information unit to be output is missing historical information is predicted as the missing probability corresponding to the information unit to be output. Based on the missing probability corresponding to the information unit to be output, each candidate information unit corresponding to the information unit to be output is determined, and the probability that the information unit to be output is each of the corresponding candidate information units is predicted to obtain the probability distribution corresponding to the information unit to be output.

[0169] Based on the probability distribution corresponding to each information unit to be output and the annotation information of the first dialogue sample, the parameters of the dialogue processing model are updated.

[0170] The model training module determines candidate information units corresponding to the information unit to be output based on the missing probability corresponding to the information unit to be output, including:

[0171] If the missing probability corresponding to the information unit to be output is greater than or equal to the preset first probability threshold, then each information unit in the second dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output.

[0172] If the missing probability corresponding to the information unit to be output is less than or equal to the preset second probability threshold, then each information unit in the first dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output.

[0173] If the missing probability corresponding to the information unit to be output is greater than the second probability threshold and less than the first probability threshold, then each word in the dictionary is determined to be a candidate information unit corresponding to the information unit to be output.

[0174] The dialogue information processing apparatus provided in this application first acquires first dialogue information (e.g., current round dialogue information) and second dialogue information (e.g., historical round dialogue information). Then, based on the first and second dialogue information, it detects whether there is missing information in the first dialogue information. When missing information is detected, the missing information is retrieved from the second dialogue information to complete the first dialogue information. The dialogue information processing apparatus provided in this application can detect whether there is missing information in the first dialogue information and, when such missing information is detected, can complete the first dialogue information. This allows for more accurate intent recognition results based on the completed dialogue information, leading to correct feedback and a better user experience.

[0175] Fifth Embodiment

[0176] This application also provides a dialogue information processing device; please refer to [link to relevant documentation]. Figure 5 The diagram shows the structure of the dialogue information processing device, which may include: at least one processor 501, at least one communication interface 502, at least one memory 503 and at least one communication bus 504.

[0177] In this embodiment of the application, the number of processor 501, communication interface 502, memory 503 and communication bus 504 is at least one, and processor 501, communication interface 502 and memory 503 communicate with each other through communication bus 504.

[0178] The processor 501 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0179] The memory 503 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0180] The memory stores a program, which the processor can call. The program is used for:

[0181] Obtain first dialogue information and second dialogue information, wherein the first dialogue information is non-first round dialogue information in multi-round dialogue information, and the second dialogue information is at least one round dialogue information that precedes the first dialogue information in the multi-round dialogue information;

[0182] Based on the first dialogue information and the second dialogue information, detect whether there is any missing information in the first dialogue information;

[0183] When missing information is detected in the first dialogue information, the missing information is obtained from the second dialogue information to complete the first dialogue information.

[0184] Optionally, the refined and extended functions of the program can be found in the description above.

[0185] Sixth Embodiment

[0186] This application embodiment also provides a readable storage medium that can store a program suitable for execution by a processor, the program being used for:

[0187] Obtain first dialogue information and second dialogue information, wherein the first dialogue information is non-first round dialogue information in multi-round dialogue information, and the second dialogue information is at least one round dialogue information that precedes the first dialogue information in the multi-round dialogue information;

[0188] Based on the first dialogue information and the second dialogue information, detect whether there is any missing information in the first dialogue information;

[0189] When missing information is detected in the first dialogue information, the missing information is obtained from the second dialogue information to complete the first dialogue information.

[0190] Optionally, the refined and extended functions of the program can be found in the description above.

[0191] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0192] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0193] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for processing dialogue information, characterized in that, include: Obtain first dialogue information and second dialogue information, wherein the first dialogue information is non-first round dialogue information in multi-round dialogue information, and the second dialogue information is at least one round dialogue information that precedes the first dialogue information in the multi-round dialogue information; Based on the first dialogue information and the second dialogue information, detect whether there is any missing information in the first dialogue information; When missing information is detected in the first dialogue information, the missing information in the first dialogue information is obtained from the second dialogue information to complete the first dialogue information; The step of detecting whether there is missing information in the first dialogue information based on the first dialogue information and the second dialogue information includes: Based on the first dialogue information and the second dialogue information, predict the probability that each information unit in the first dialogue information is missing historical context information; Based on the predicted probability, determine whether each information unit in the first dialogue information is missing historical context.

2. The dialogue information processing method according to claim 1, characterized in that, The step of detecting whether there is missing information in the first dialogue information based on the first dialogue information and the second dialogue information, and, when missing information is detected in the first dialogue information, obtaining the missing information from the second dialogue information to complete the first dialogue information, includes: The first dialogue information and the second dialogue information are processed using a pre-trained dialogue processing model to obtain the completed dialogue information corresponding to the first dialogue information. The dialogue processing model has the ability to detect and complete missing information in the first dialogue information.

3. The dialogue information processing method according to claim 2, characterized in that, The step of processing the first dialogue information and the second dialogue information using a pre-trained dialogue processing model to obtain the completed dialogue information corresponding to the first dialogue information includes: The dialogue processing model is used to process the first dialogue information and the second dialogue information to determine and output the information units to be output one by one. The process of determining a unit of information to be output using the dialogue processing model includes: The probability that the information unit to be output is missing historical context is predicted using the dialogue processing model, and the information unit to be output is determined based on the predicted probability.

4. The dialogue information processing method according to claim 3, characterized in that, The step of determining the information unit to be output based on the predicted probability includes: If the predicted probability is greater than or equal to the preset probability threshold, then select an information unit from the second dialogue information as the information unit to be output. If the predicted probability is less than the probability threshold, then an information unit is obtained from the first dialogue information and used as the information unit to be output.

5. The dialogue information processing method according to claim 3, characterized in that, The step of determining the information unit to be output based on the probability that the information unit to be output is missing historical information includes: If the predicted probability is greater than or equal to the preset first probability threshold, then select an information unit from the second dialogue information as the information unit to be output. If the predicted probability is less than or equal to the preset second probability threshold, then the information unit is obtained from the first dialogue information and used as the information unit to be output. If the predicted probability is greater than the second probability threshold and less than the first probability threshold, then an information unit is generated based on the output information unit and the dictionary, and used as the information unit to be output.

6. The dialogue information processing method according to claim 2, characterized in that, The dialogue processing model includes an encoding layer and a decoding layer; The encoding layer encodes the embedded representation of the first dialogue information and the embedded representation of the second dialogue information to obtain the encoding results corresponding to the first dialogue information and the second dialogue information respectively; At each decoding moment, the decoding layer predicts the probability that the information unit to be output at the current decoding moment is missing historical context information, based on the encoding results corresponding to the first dialogue information and the second dialogue information, respectively, and the information unit prediction vector of the previous decoding moment. The layer then determines the information unit to be output based on the predicted probability and outputs it.

7. The dialogue information processing method according to claim 6, characterized in that, The process of obtaining the embedded representation of dialogue information includes: Obtain the representation vector of each information unit in the dialogue information, the position representation vector of each information unit in the dialogue information, and the dialogue round representation vector of the dialogue information; The representation vector of each information unit in the dialogue information, the position representation vector of each information unit in the dialogue information, and the dialogue turn number representation vector of the dialogue information are fused together, and the fused vector is used as the embedded representation of the dialogue information.

8. The dialogue information processing method according to claim 6, characterized in that, The decoding layer includes: a first attention module, a second attention module, and a missing information detection and completion module; The first attention module determines the information unit prediction vector for the current decoding time based on the information unit prediction vector from the previous decoding time. The second attention module determines a first vector that can characterize the correlation between the information unit to be output at the current decoding time and each information unit in the first dialogue information, and a second vector that can characterize the correlation between the information unit to be output at the current decoding time and each information unit in the second dialogue information, based on the encoding results corresponding to the first dialogue information and the information unit prediction vector at the current decoding time. The missing information detection and completion module predicts the probability that the information unit to be output at the current decoding moment is missing historical information based on the first vector and the second vector, and outputs the information unit according to the predicted probability.

9. The dialogue information processing method according to claim 8, characterized in that, The second attention module includes: an attention module corresponding to the first dialogue information and an attention module corresponding to the second dialogue information; The attention module corresponding to the first dialogue information determines the weight of each information unit in the first dialogue information based on the information unit prediction vector at the current decoding time and the encoding result corresponding to the first dialogue information, and determines the first vector based on the determined weight and the encoding result corresponding to the first dialogue information. The attention module corresponding to the second dialogue information determines the weight of each information unit in the second dialogue information based on the information unit prediction vector at the current decoding time and the encoding result corresponding to the second dialogue information, and determines the second vector based on the determined weight and the encoding result corresponding to the second dialogue information.

10. The dialogue information processing method according to claim 2, characterized in that, The training process of the dialogue processing model includes: Training data is obtained from the training dataset, wherein the training data includes a first dialogue sample and a second dialogue sample, and the first dialogue sample is labeled with dialogue information without missing information corresponding to the first dialogue sample. The first dialogue sample and the second dialogue information are processed using a dialogue processing model to predict the probability distribution corresponding to each information unit to be output. When predicting the probability distribution corresponding to an information unit to be output, the probability that the information unit to be output is missing historical information is predicted as the missing probability corresponding to the information unit to be output. Based on the missing probability corresponding to the information unit to be output, each candidate information unit corresponding to the information unit to be output is determined, and the probability that the information unit to be output is each of the corresponding candidate information units is predicted to obtain the probability distribution corresponding to the information unit to be output. Based on the probability distribution corresponding to each information unit to be output and the annotation information of the first dialogue sample, the parameters of the dialogue processing model are updated.

11. The dialogue information processing method according to claim 10, characterized in that, The step of determining each candidate information unit corresponding to the information unit to be output based on the missing probability of the information unit to be output includes: If the missing probability corresponding to the information unit to be output is greater than or equal to the preset first probability threshold, then each information unit in the second dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output. If the missing probability corresponding to the information unit to be output is less than or equal to the preset second probability threshold, then each information unit in the first dialogue sample is determined to be a candidate information unit corresponding to the information unit to be output. If the missing probability corresponding to the information unit to be output is greater than the second probability threshold and less than the first probability threshold, then each word in the dictionary is determined to be a candidate information unit corresponding to the information unit to be output.

12. A dialogue information processing device, characterized in that, include: The module includes a dialogue information acquisition module, an information missing detection module, and a dialogue information completion module. The dialogue information acquisition module is used to acquire first dialogue information and second dialogue information, wherein the first dialogue information is non-first round dialogue information in multi-round dialogue information, and the second dialogue information is at least one round of dialogue information that precedes the first dialogue information in the multi-round dialogue information. The information missing detection module is used to detect whether there is any missing information in the first dialogue information based on the first dialogue information and the second dialogue information. The dialogue information completion module is used to obtain the missing information in the first dialogue information from the second dialogue information and complete the first dialogue information when the information missing detection module detects that there is missing information in the first dialogue information. Specifically, the information missing detection module is used to predict the probability that each information unit in the first dialogue information is missing historical context information based on the first dialogue information and the second dialogue information; and to determine whether each information unit in the first dialogue information is missing historical context information based on the predicted probability.

13. A dialogue information processing device, characterized in that, include: Memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement each step of the dialogue information processing method as described in any one of claims 1 to 11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the dialogue information processing method as described in any one of claims 1 to 11.

Images