Methods for training information generation models, methods for generating information, apparatus, electronic devices, storage media, and computer programs.

Abstract

To provide a training method for an information generation model capable of generating recommendation information, an information generation method, an apparatus, an electronic device, a storage medium, and a computer program.SOLUTION: The training method for an information generation model divides descriptive information for a target object in an information pair into at least one descriptive word to obtain a descriptive word sequence. The information pair further includes first recommendation information. The method further includes inputting the descriptive word sequence into a dialog generation model to obtain a probability vector sequence for the target object. Each probability vector in the probability vector sequence includes probability values for a plurality of predetermined words. The method further includes training an interaction generative model and obtaining an information generative model based on the probability vector sequence and the first recommendation information.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] This disclosure relates to the field of artificial intelligence, specifically to the technologies of natural language processing and deep learning, and in particular to methods for training information generation models, methods for generating information, apparatus, electronic devices, storage media, and computer programs. [Background technology]

[0002] With the advancement of computer and network technologies, methods for generating information using natural language processing techniques have become widespread. To contribute to the dissemination of information, it is necessary to consider the consistency and non-redundancy of the information. [Overview of the project] [Means for solving the problem]

[0003] This disclosure aims to provide a method for training an information generation model, an information generation method, an apparatus, an electronic storage medium, and a computer program that improve the diversity of the generated information.

[0004] According to one aspect of the present disclosure, a method for training an information generation model is provided, the method comprising: dividing descriptive information for a target object in an information pair into at least one descriptive word to obtain a descriptive word sequence, wherein the information pair further includes first recommendation information; inputting the descriptive word sequence into a dialogue generation model to obtain a probability vector sequence for a target object, wherein each probability vector in the probability vector sequence includes probability values ​​for a plurality of predetermined words; and training the dialogue generation model and obtaining an information generation model based on the probability vector sequence and the first recommendation information.

[0005] According to one aspect of the present disclosure, an information generation method is provided, which includes: dividing descriptive information of an object to be recommended into at least one descriptive word to obtain a descriptive word sequence; inputting the descriptive word sequence into an information generation model to obtain a probability vector sequence for the object to be recommended, wherein each probability vector in the probability vector sequence contains probability values ​​for a plurality of predetermined words; and identifying recommendation information for the object to be recommended based on the probability vector sequence, wherein the information generation model is trained using an information generation model training method provided in the present disclosure.

[0006] According to one aspect of the present disclosure, a training device for an information generation model is provided, the device comprising: a first splitting module for obtaining a descriptive word sequence by splitting descriptive information for a target object in an information pair into at least one descriptive word, wherein the information pair further includes first recommendation information; a sequence acquisition module for inputting the descriptive word sequence into a dialogue generation model to obtain a probability vector sequence for a target object, wherein each probability vector in the probability vector sequence includes probability values ​​for a plurality of predetermined words; and a training module for training a dialogue generation model and obtaining an information generation model based on the probability vector sequence and the first recommendation information.

[0007] According to one aspect of the present disclosure, an information generation device is provided, the device comprising: a first division module for dividing descriptive information of an object to be recommended into at least one descriptive word to obtain a descriptive word sequence; a sequence acquisition module for inputting the descriptive word sequence into an information generation model to obtain a probability vector sequence for an object to be recommended, wherein each probability vector in the probability vector sequence includes probability values ​​for a plurality of predetermined words; and an information identification module for identifying recommendation information for an object to be recommended based on the probability vector sequence, the information generation model is trained using an information generation model training device provided in the present disclosure.

[0008] According to another aspect of the present disclosure, an electronic device is provided, the electronic device comprising at least one processor and a memory communicated with the at least one processor, the memory storing commands to be executed by the at least one processor, the execution of the commands by the at least one processor enabling the at least one processor to execute a method for training an information generation model and / or an information generation method provided by the present disclosure.

[0009] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided which stores computer commands for causing a computer to perform a training method for an information generation model and / or an information generation method provided by this disclosure.

[0010] According to another aspect of this disclosure, a computer program is provided which, when executed on a processor, realizes a method for training an information generation model and / or an information generation method provided by this disclosure.

[0011] It should be understood that the contents described in this section are not intended to highlight any essential or important features of the embodiments of this disclosure, nor to limit the scope of this disclosure. Other features of this disclosure will be readily apparent throughout the following specification.

[0012] The drawings are provided to better understand this solution and do not limit the scope of this disclosure. [Brief explanation of the drawing]

[0013] [Figure 1] Figure 1 is a schematic diagram illustrating application scenarios for the training method and information generation method and apparatus of the information generation model according to the embodiments of this disclosure. [Figure 2] Figure 2 is a schematic flowchart of the training method for the information generation model according to an embodiment of the present disclosure. [Figure 3] Figure 3 is a schematic diagram illustrating the principle of the training method for an information generation model according to an embodiment of this disclosure. [Figure 4] Figure 4 is a schematic diagram illustrating the principle of a training method for an information generation model according to another embodiment of this disclosure. [Figure 5] Figure 5 is a schematic diagram illustrating the principle of a training method for an information generation model according to another embodiment of the present disclosure. [Figure 6] Figure 6 is a schematic flowchart of the information generation method according to an embodiment of the present disclosure. [Figure 7] Figure 7 is a block diagram of the configuration of the training device for the information generation model according to an embodiment of the present disclosure. [Figure 8] Figure 8 is a block diagram of the configuration of an information generation device according to an embodiment of the present disclosure. [Figure 9] Figure 9 is a block diagram of an electronic device for implementing the training method and / or information generation method of the information generation model according to an embodiment of the present disclosure. [Modes for carrying out the invention]

[0014] Illustrative embodiments of the present disclosure will be described below with reference to the drawings, and for ease of understanding, they will include various details of the embodiments of the present disclosure, which should be considered merely illustrative. Accordingly, those skilled in the art should understand that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for clarity and conciseness, descriptions of well-known functions and configurations will be omitted in the following description.

[0015] The present disclosure provides a method for training an information generation model, the method including a splitting stage, a sequence acquisition stage, and a training stage. In the splitting stage, the description information for the target object in the information pair is split into at least one description word to obtain a description word sequence, where the information pair further includes first recommendation information. In the sequence acquisition stage, the description word sequence is input into a dialogue generation model to obtain a probability vector sequence for the target object, and each probability vector in the probability vector sequence includes probability values for a plurality of predetermined words. In the training stage, based on the probability vector sequence and the first recommendation information, the dialogue generation model is trained to obtain an information generation model.

[0016] Hereinafter, with reference to FIG. 1, the application scenarios of the method and apparatus provided by the present disclosure will be described.

[0017] FIG. 1 is a schematic diagram of an application scenario of a method for training an information generation model and an information generation method and apparatus according to an embodiment of the present disclosure.

[0018] As shown in FIG. 1, the application scenario 100 of this embodiment may include an electronic device 110, and the electronic device 110 may be various electronic devices having a processing function, and may include, but is not limited to, a smartphone, a tablet computer, a laptop portable computer, a desktop computer, a server, etc.

[0019] According to an embodiment of the present disclosure, the electronic device 110 may, for example, provide a man-machine interface and be used to obtain description information 120 through a user's operation. This description information 120 may, for example, be for an item to be promoted. The electronic device 110 may, for example, generate recommendation information 130 for the item to be promoted based on the obtained description information 120.

[0020] For example, the electronic device 110 may, for instance, employ text summarization technology to extract key words or key fragments from the descriptive information 120, and then arrange and combine the extracted key words or key fragments to obtain recommendation information 130.

[0021] For example, an electronic device may employ an end-to-end information generation model, input descriptive information 120 into the information generation model, and output recommendation information 130 through the information generation model. Here, the end-to-end information generation model may be obtained by training with a collected advertising corpus. Here, the information generation model may employ a recurrent neural network (RNN), and more specifically, a bidirectional recurrent neural network (BiRNN), a gated recurrent unit (GRU), or a long short-term memory (LSTM). Alternatively, the information generation model may employ a Transformer model or the like.

[0022] According to embodiments of the present disclosure, the application scenario 100 may further include a server 150, as shown in Figure 1. The server 150 may train an information generation model based on a corpus and, in response to acquisition requests transmitted from the electronic device 110, transmit the trained information generation model 140 to the electronic device 110, thereby enabling the electronic device 110 to generate recommendation information for items to be promoted using the information generation model 140. To understand this, the electronic device 110 may further transmit acquired descriptive information 120 to the server 150, which may then generate recommendation information for items to be promoted using the information generation model 140.

[0023] To make it easier to understand, the electronic device 110 may communicate with the server 150 via a network, and the network may be a wired or wireless link. The server 150 may be a server that provides various services, for example, a background management server (this is just an example) that provides support for the human-machine interface run by the electronic device 110. The server 150 may be a server in a distributed system, or a server that incorporates a blockchain.

[0024] It should be explained that the information generation method provided in this disclosure may be executed by the electronic device 110 or by the server 150. Accordingly, the information generation device provided in this disclosure may be installed on the electronic device 110 or by the server 150. The information generation model training method provided in this disclosure may be executed by the server 150. Accordingly, the information generation model training device provided in this disclosure may be installed on the server 150.

[0025] It should be understood that the number and types of electronic devices 110 and servers 150 in Figure 1 are merely illustrative. Any number and types of electronic devices 110 and servers 150 may be used as required by the implementation.

[0026] The training method for the information generation model provided in this disclosure will be explained in detail below, with reference to Figure 1 and Figures 2 to 5.

[0027] Figure 2 is a schematic flowchart of the training method for the information generation model according to an embodiment of the present disclosure.

[0028] As shown in Figure 2, the training method 200 for the information generation model in this embodiment may include operations S210 to S230.

[0029] In operation S210, the descriptive information for the target object in the information pair is divided into at least one descriptive word to obtain a descriptive word sequence.

[0030] According to embodiments of this disclosure, an information pair may include descriptive information and recommendation information. For example, the descriptive information and recommendation information may be for the same target, and thus the information pair can be a positive sample. For example, the descriptive information and recommendation information may be for different target, and thus the information pair can be a negative sample. For example, there may be multiple information pairs, and multiple information pairs may simultaneously include positive and negative samples. Here, the target may be any items to be promoted, such as clothing, food, or furniture.

[0031] According to embodiments of this disclosure, descriptive information may be divided using a word segmentation tool to obtain a descriptive word sequence. Here, the word segmentation tool may be the jieba word segmentation tool, the jiagu word segmentation tool, or the NLPIR word segmentation tool, etc., and this disclosure is not limited thereto. For example, if the descriptive information is "children's down jacket", a word sequence consisting of the words "children" and "down jacket" can be obtained by word segmentation, and this word sequence is a descriptive word sequence.

[0032] To ensure understanding, the first recommendation information may be, but is not limited to, a promoted target recommendation or a manually set recommendation, and the first recommendation information may specifically be a promoted target recommendation ad word.

[0033] In operation S220, the descriptive word sequence is input to the dialogue generation model to obtain a probability vector sequence for the target object.

[0034] According to embodiments of this disclosure, the dialogue generation model may include, for example, an encoding network and a decoding network. Here, the encoding network and the decoding network may be selected from a Convolutional Neural Network (CNN), an RNN, or a self-attention network. Here, the RNN may include BiRNN, GRU, or LSTM, etc. The embodiment may directly input a descriptive word sequence into the dialogue generation model and output a probability vector sequence by the dialogue generation model. Alternatively, the embodiment may first embed each word in the descriptive word sequence and obtain a word feature vector representing each word. The word feature vector may also be a word identifier (token) representing a word, in which case a token sequence can be obtained after embedding the word sequence. The token sequence can be input into the dialogue generation model to obtain a probability vector sequence. It should be noted that each probability vector in the probability vector sequence is output sequentially, and multiple tokens in the token sequence may be input into the dialogue generation model simultaneously.

[0035] In one embodiment, the dialogue generation model may employ a pre-trained dialogue generation model with discrete latent variables, which can effectively model "one-to-many" relationships in dialogues and help improve the diversity of the generated recommendation information. Here, the pre-trained dialogue generation model with discrete latent variables may be a model from the "Asphalt" (PLATO) series, such as the PLATO-2 model, and is not limited to this disclosure.

[0036] According to embodiments of this disclosure, the obtained probability vector sequence may include a plurality of probability vectors, each probability vector may include probability values ​​for a plurality of predetermined words. Each probability vector may be a token representing a single recommended word, and the plurality of predetermined words may be words included in a predefined dictionary.

[0037] In operation S230, a dialogue generation model is trained and an information generation model is obtained based on a probability vector sequence and the first recommendation information.

[0038] According to embodiments of this disclosure, first, the first recommendation information may be divided into at least one recommendation word to obtain a recommendation word sequence. Next, based on a predefined dictionary, an embedded representation of each recommendation word in the recommendation word sequence may be obtained, and a word vector representing each recommendation word may be obtained, thus converting the recommendation word sequence into a word vector sequence. The embodiment may identify the loss of the dialogue generation model based on the distance between each word vector in the word vector sequence and the probability vector corresponding to each word vector in the probability vector sequence. By adjusting the network parameters in the dialogue generation model using a backward propagation algorithm, the loss of the dialogue generation model is minimized and the dialogue generation model reaches a predetermined convergence condition. In this embodiment, the dialogue generation model that has reached the predetermined convergence condition may be used as the information generation model.

[0039] In one embodiment, the dialogue generation model may be trained by maximizing the probability of generating the first recommendation information. For example, after obtaining at least one recommendation word, the t-th recommendation word y in the t-th probability vector of the probability vector sequence is obtained from the at least one recommendation word. t The probability value P(y) t |x,y <t ) may be identified. Here, if both the number of recommended words and the number of probability vectors in the probability vector sequence are set to T, the loss of the dialogue generation model may be identified using the negative log-likelihood loss (NLL loss). Specifically, the loss L of the dialogue generation model NLL This can also be obtained by calculating it using the following formula (1).

number

[0040] Here, S represents the training sample set, and y <t x represents the information output by the dialogue generation model before it outputs the t-th probability vector, and x represents the input information for the dialogue generation model.

[0041] As can be seen from the above, the method for training an information generation model in the embodiment of this disclosure acquires an information generation model that generates recommendation information by training a dialogue generation model, effectively increases the corpus of the information generation model acquired through training, helps to improve the diversity and consistency of the recommendation information acquired using the information generation model, and helps to avoid to some extent situations where the generated recommendation information does not conform to common sense and is too heavily marketed, thereby helping to improve the effectiveness of the target promotion. For example, the dialogue generation model adopted in this embodiment may be a model obtained by training in advance using a natural dialogue corpus, and in this way the obtained recommendation generation model has learned a richer amount of common sense knowledge.

[0042] In one embodiment, when training a dialogue generation model, the dialogue generation model may first be pre-trained using a large amount of natural dialogue data. The principle of this pre-training is similar to the implementation principle of method 200 described above, and will not be explained here. After pre-training, the embodiment may train the pre-trained dialogue generation model using a question-and-answer corpus of the information recommendation field as training data. In this way, the dialogue generation model can be transitioned from the natural field to the recommendation field through learning. Accordingly, the dialogue generation model in operation S220 described above may be a dialogue generation model obtained after first undergoing pre-training and then training using a question-and-answer corpus of the information recommendation field.

[0043] Figure 3 is a schematic diagram illustrating the principle of the training method for an information generation model according to an embodiment of this disclosure.

[0044] According to embodiments of this disclosure, presentation information can be provided to an information generation model so that the resulting recommendation information better matches actual needs and better embodies the highlights of the products to be promoted. Here, the presentation information and the aforementioned descriptive information may both be provided, for example, by the manufacturer or seller of the target product.

[0045] Specifically, as shown in Figure 3, in the embodiment 300, after obtaining descriptive information 310 and presentation information 320 by a human-machine interface, the word segmentation tool described in the preceding paragraph may be used to divide the descriptive information 310 into at least one descriptive word to obtain a descriptive word sequence 311, and at the same time, the presentation information 320 may be divided into at least one presentation word to obtain a presentation word sequence 321. Next, both the descriptive word sequence 311 and the presentation word sequence 321 may be input to the described dialogue generation model 301, and the dialogue generation model 301 may output a probability vector sequence. For example, if the descriptive information is "children's down jacket", the presentation information may be "casual, warm, free shipping", and accordingly, the presentation word sequence 321 may be a word sequence composed of the word "casual", the word "warm", and the word "free shipping".

[0046] According to embodiments of this disclosure, after obtaining a probability vector sequence 330, the predicted probability that the second recommendation information includes the presented information 320 may be identified based on the probability vector sequence 330. Then, with the training objective being that the second recommendation information includes the presented information 320, a dialogue generation model is trained based on the predicted probability that the second recommendation information includes the presented information 320. In this way, the probability that the generated recommendation information includes the presented information can be improved, and the generated recommendation information will better match the actual needs.

[0047] In one embodiment, the second recommendation information described above may first be identified based on the probability vector sequence 330. Specifically, for each probability vector, a predetermined word corresponding to the maximum probability value in each probability vector may be identified, and one predicted word may be obtained. The obtained multiple predicted words are sorted based on their positions in the probability vector sequence 330, and the multiple predicted words are combined to constitute the second recommendation information. Next, the second recommendation information is searched using each word in the presented word sequence, and it is determined whether the second recommendation information contains each word. The proportion of the second recommendation information that contains a word in the presented word sequence is taken as the predicted probability that the second recommendation information contains the presented information.

[0048] For example, if the multiple predicted words obtained by sorting are the words "such as", "child", "down jacket", "comfortable", "warm", "A", "region", and "free shipping", then the second recommendation information could be "Such children's down jackets are comfortable and warm, and shipping is free to region A".

[0049] In one embodiment, for each suggested word in the suggested word sequence 321, the probability that the second recommendation information includes each suggested word may be identified based on the probability values ​​of all probability vectors in the probability vector sequence for each suggested word. Specifically, at least one probability value for each suggested word may be obtained based on all probability vectors, and in this embodiment, the average or minimum value of this at least one probability value may be taken as the probability that the second recommendation information includes each suggested word. For at least one suggested word included in the suggested word sequence 321, at least one probability 340 may be obtained that the second recommendation information includes at least one suggested word. In this embodiment, a predicted probability 350 that the second recommendation information includes suggested information 320 may be identified based on this at least one probability 340. For example, the sum of these at least one probabilities 340 may be taken as the predicted probability 350.

[0050] After obtaining the prediction probability 350 that the second recommendation information includes the presentation information, the embodiment may specify the first loss value 360 of the dialogue generation model 301 based on the first probability. Next, based on this first loss value 360, the dialogue generation model 301 is trained. For example, the first loss value 360 may be specified using a predetermined loss function, where the independent variable in the predetermined loss function is the prediction probability 350, the dependent variable is the first loss value 360, and the first loss value 360 has a negative correlation with the prediction probability 350, thereby achieving the training objective that the second recommendation information includes the presentation information 320.

[0051] In one embodiment, the predetermined loss function for specifying the first loss value 360 may be represented by the following formula (2).

Number

[0052] Here, k j represents the j-th presentation word in at least one presentation word, p i (k j |x,y <t ) represents the probability value for the j-th presentation word of the i-th probability vector in the probability vector sequence, N is the total number of at least one presentation word, T is the total number of probability vectors included in the probability vector sequence, and L k is the first loss value.

[0053] The embodiments of the present disclosure can achieve the effect that the words generated by the dialogue generation model are controllable to a certain extent by inputting the presentation information into the dialogue generation model and training the dialogue generation model based on the prediction probability that the second recommendation information includes the presentation information. In this way, when generating recommendation information using the information generation model obtained by training, the presentation information that is desired to appear in the recommendation information can be included.

[0054] FIG. 4 is a schematic diagram of the principle of a method for training an information generation model according to another embodiment of the present disclosure.

[0055] According to embodiments of this disclosure, a dialogue generation model may be trained based on the relationship between second recommendation information and descriptive information indicated by a probability vector sequence, thereby bringing the themes of the second recommendation information closer to the themes described in the descriptive information, and thereby improving, to some extent, the accuracy of the recommendation information generated using the trained information generation model.

[0056] In one embodiment, the embodiment may identify a second recommendation information based on a probability vector sequence by employing the method described above. Next, it identifies the relationship between the second recommendation information and the descriptive information, and identifies a second loss value for the dialogue generation model based on the relationship. Finally, it trains the dialogue generation model based on the second loss value.

[0057] For example, the relationship between the second recommendation information and the descriptive information can be shown using semantic similarity. For instance, a semantic feature extraction model may be used to extract semantic features from the second recommendation information and the descriptive information, and the similarity between the two semantic features may be defined as semantic similarity. The semantic feature extraction model may be an LSTM model or the like, and this disclosure is not limited thereto. Alternatively, the relationship between the second recommendation information and the descriptive information can be shown using thematic similarity. Here, a theme extraction model may be used to extract the themes from the second recommendation information and the descriptive information, and the similarity between the two themes may be defined as thematic similarity. The theme extraction model may be a latent Dirichlet allocation (LDA) model or the like, and this disclosure is not limited thereto. Accordingly, the embodiment may identify the second loss value using a predetermined loss function that has a negative correlation with the semantic similarity or thematic similarity.

[0058] In one embodiment, the dialogue generation model may employ a pre-trained dialogue generation model with discrete latent variables. As shown in Figure 4, the embodiment 400 enables the dialogue generation model to perform multi-task learning by adding an auxiliary task to the dialogue generation model. Specifically, one identification bit 410 may be added to the input of the dialogue generation model, i.e., one token may be increased, and the identification bit 410 may employ arbitrary random identification information. As can be understood, the identification bit is similar in function to the latent variable bits input to the PLATO-2 model described above, and this disclosure omits further explanation. When training the dialogue generation model 401, the embodiment may input the random identification information and descriptive word sequence 420 adopted by the identification bit 410 to the dialogue generation model 401, and the dialogue generation model may output association prediction values ​​430 and probability vector sequence 440 corresponding to the random identification information. Here, the association prediction values ​​430 are used to indicate the association relationship between second recommendation information and descriptive information.

[0059] In this embodiment, the dialogue generation model may be trained with the goal of maximizing this related prediction value. That is, the second loss has a negative correlation with the related prediction value. For example, the dialogue generation model 401 may first identify a second loss value 450 using a predetermined loss function based on the related prediction value 430, and then train the dialogue generation model 401 with the goal of minimizing the second loss value 450. At the same time, the embodiment may further adopt equation (1) described above and identify a negative log-likelihood loss value 460 based on the probability vector sequence 440. The embodiment may identify a total loss value 470 using a weighted sum of the second loss value 450 and the negative log-likelihood loss value 460, and then train the dialogue generation model 401 based on the total loss value.

[0060] For example, the predetermined loss function that identifies the second loss value of 450 may be a cross-entropy loss function, and this disclosure is not limited thereto.

[0061] In one embodiment, a second loss value of 450 may be considered simultaneously with the aforementioned first loss value, so that the input to the dialogue generation model includes not only random identification information and descriptive word sequences, but also presentation word sequences. In this embodiment, the total loss value can be determined by a weighted sum of the negative log-likelihood loss value of 460, the second loss value of 450, and the aforementioned first loss value. As can be understood, the weight coefficients used when calculating the weighted sum may be set as needed and are not limited to those.

[0062] In this embodiment, a pre-trained dialogue generation model with discrete latent variables is employed as the dialogue generation model, and random identification information is added to the input of the dialogue generation model, enabling the dialogue generation model to complete multi-task learning. The associated predicted values ​​corresponding to the output identification bits can provide guidance to some extent, specifically, when generating recommendation information using the information generation model obtained through training, they can provide guidance on whether or not the recommendation information is adopted.

[0063] Figure 5 is a schematic diagram illustrating the principle of a training method for an information generation model according to another embodiment of the present disclosure.

[0064] According to embodiments of this disclosure, the dialogue generation model may be trained with the goal of ensuring that the generated recommendation information does not contain duplicate words. This results in recommendation information generated using the trained information generation model being more concise and smoother, which is useful for promoting the target to be recommended in relation to the recommendation information.

[0065] In one embodiment, a second recommendation for the target object may be identified based on a probability vector sequence using a method similar to the one described above. Next, it may be determined whether or not there are duplicate words among the multiple predicted words obtained in the process of identifying the second recommendation information. If duplicate words exist, the probability vector corresponding to the duplicate word in the probability vector sequence may be identified based on the position information of the duplicate word in the second recommendation information, and this probability vector may be set as the target probability vector. Specifically, for example, the multiple predicted words are A, B, C, A, and D in order, and each of these multiple predicted words corresponds to the word with the highest probability value included in the five probability vectors in the probability vector sequence. Since the first word among the multiple predicted words overlaps with the fourth word, the embodiment may set the first and fourth probability vectors among the five probability vectors as the target probability vectors. Subsequently, the embodiment may identify a third loss value for the dialogue generation model based on the target probability vector. For example, the embodiment may identify the third loss value based on the highest probability value included in any one of the target probability vectors. To make it clear, the maximum probability value contained in any one of the probability vectors is the probability of the duplicate word contained in that one probability vector, i.e., a third loss value may be identified based on the target probability vector and the duplicate word.

[0066] In one embodiment, the target probability vector in the probability vector sequence may be identified based on the duplicate words contained in the first recommendation information in the information pair. This is because the training objective of the dialogue generation model is to match the second recommendation information with the first recommendation information. Identifying the target probability vector based on the duplicate words contained in the first recommendation information can improve training efficiency.

[0067] Specifically, as shown in Figure 5, in the embodiment 500, when training the dialogue generation model 501, the descriptive word sequence W_d 510 is input to the dialogue generation model 501 to obtain the probability vector sequence V_p 520, and at the same time, the first recommendation information is divided into at least one recommendation word to obtain the recommendation word sequence W_t 530. Here, the descriptive word sequence W_d 510 may contain a total of M descriptive words, from W_d1 to W_dM. The recommendation word sequence W_t 530 may contain a total of T recommendation words, from W_d1 to W_dT. The probability vector sequence V_p 520 contains a total of T probability vectors, from V_p1 to V_pT.

[0068] After obtaining the recommended word sequence W_t 530, the embodiment may determine whether or not there are duplicate words in the recommended word sequence. If there are no duplicate words, the third loss value is determined to be 0. If there are duplicate words, the embodiment may determine the probability vectors corresponding to the duplicate words in the probability vector sequence based on the positional information of the duplicate words in the actual word sequence. For example, if recommended words W_d1 and W_d3 are duplicate words in the actual word sequence, the probability vectors corresponding to those duplicate words in the probability vector sequence V_p 520 are probability vectors V_p1 and V_p3, i.e., the target probability vectors include probability vectors V_p1 and V_p3. After determining the target probability vectors, the embodiment may determine a third loss value for the dialogue generation model based on the target probability vectors and the duplicate words. For example, the third loss value may be determined based on the probability value for the duplicate words contained in any one of the target probability vectors. For example, a target probability vector may be defined as the probability vector corresponding to the word that is closer in position among the duplicate words, and a third loss value may be identified based on the probability values ​​for the duplicate words included in the target probability vector.

[0069] According to embodiments of this disclosure, the third loss value may have a positive correlation with the probability value for duplicate words included in the target probability vector. This allows the dialogue generation model to be trained based on the third loss value to minimize the probability value for duplicate words in the target probability vector and reduce the probability of duplicate words appearing. To understand this, if there are more than one type of duplicate word in the second recommendation information, a set of target probability vectors may be identified for each type of duplicate word, and the embodiment may identify one loss value based on each set of target probability vectors, with the sum of the identified loss values ​​being the third loss value.

[0070] In one embodiment, the third loss value L rep This can be shown by the following equation (3).

number

[0071] Here, C is a set of duplicate words composed of duplicate words, c is one of the duplicate words in the set of duplicate words, x is the input information of the dialogue generation model, and y <t This indicates the information output before the dialogue generation model outputs the t-th probability vector (i.e., the target probability vector), and p(c|x,y <t ) indicates the probability value for the duplicate word c included in the target probability vector.

[0072] To make it clearer, in this embodiment, the dialogue generation model may be trained using the weighted sum of the third loss value and the negative log-likelihood loss value obtained by equation (1) above as the total loss value.

[0073] To make it clear, the embodiment may determine a total loss value based on the aforementioned negative log-likelihood loss value and any combination of the first to third loss values, and train a dialogue generation model based on the total loss value.

[0074] Based on an information generation model trained according to the information generation model training method provided in this disclosure, this disclosure further provides an information generation method, which will be described in detail below with reference to Figure 6.

[0075] Figure 6 is a schematic flowchart of the information generation method according to an embodiment of the present disclosure.

[0076] As shown in Figure 6, the information generation method 600 of this embodiment may include operations S610 to S630.

[0077] In operation S610, the descriptive information of the target to be recommended is divided into at least one descriptive word to obtain a descriptive word sequence.

[0078] According to the embodiments of this disclosure, the target to be recommended is similar to the target target described above, and the method for realizing operation S610 is similar to the method for realizing operation S210 described above, so an explanation is omitted here.

[0079] In operation S620, the descriptive word sequence is input into the information generation model to obtain a probability vector sequence for the target to be recommended.

[0080] Here, each probability vector in the probability vector sequence contains probability values ​​for a plurality of predetermined words. The information generation model is trained by employing the information generation model training method provided in this disclosure. Operation S620 is similar to the implementation method of operation S220 described above, and its explanation is omitted here.

[0081] In operation S630, recommendation information for the target to be recommended is identified based on a probability vector sequence.

[0082] According to embodiments of this disclosure, operation S630 may be implemented by employing a method similar to the method for identifying second recommendation information based on the aforementioned probability vector sequence, but this will not be explained here.

[0083] According to embodiments of this disclosure, the information generation method 600 may further include an operation to obtain a sequence of suggested words by dividing the suggested information for the object to be recommended into at least one suggested word. Thus, the aforementioned operation S620 may specifically input the descriptive word sequence and the suggested word sequence into a dialogue generation model to obtain a probability vector sequence. The implementation principle of this embodiment is similar to the principle by which embodiment 300 obtained a probability vector sequence, and will not be explained here.

[0084] According to embodiments of this disclosure, the dialogue generation model may include a pre-trained dialogue generation model with discrete latent variables. The aforementioned operation S620 may specifically involve inputting random identifiers and descriptive word sequences into the dialogue generation model to obtain association values ​​and probability vector sequences corresponding to the random identifiers. Here, the association values ​​indicate the association relationship between the recommendation information and the descriptive information. The implementation principle of this embodiment is similar to the principle by which Embodiment 400 obtained association prediction values ​​and probability vector sequences, and is therefore omitted from this explanation.

[0085] What needs to be explained is that the information generation method provided in this disclosure employs a model obtained by training a dialogue generation model, so it can understand the meaning of the information input to the model, and at the same time provide appropriate answers to the information with questions input to the model, it can provide common sense knowledge to the resulting recommendation information, which helps to improve the effectiveness of the promotion of the target to be recommended corresponding to the recommendation information. For example, if the information input to the model includes "What is the use of skincare ingredient XXX?", the resulting recommendation information may include "Skincare ingredient XXX can fill in wrinkles and depressions on the face and make the skin more moisturized," etc.

[0086] Based on the information generation model training method provided in this disclosure, this disclosure further provides an information generation model training apparatus, which will be described below with reference to Figure 7.

[0087] Figure 7 is a block diagram showing the configuration of a training device for an information generation model according to an embodiment of this disclosure.

[0088] As shown in Figure 7, the training device 700 for the information generation model of this embodiment includes a first division module 710, a sequence acquisition module 720, and a training module 730.

[0089] The first splitting module 710 is for splitting the descriptive information for the target object in the information pair into at least one descriptive word to obtain a descriptive word sequence, where the information pair further includes first recommendation information. In one embodiment, the first splitting module 710 may also be for performing the operation S210 described above, which will not be explained here.

[0090] The sequence acquisition module 720 is used to input a descriptive word sequence into a dialogue generation model and acquire a probability vector sequence for the target object. Here, each probability vector in the probability vector sequence contains probability values ​​for multiple predetermined words. In one embodiment, the sequence acquisition module 720 may also be used to perform the operation S220 described above, and this will not be explained here.

[0091] The training module 730 is for training a dialogue generation model and obtaining an information generation model based on a probability vector sequence and first recommendation information. In one embodiment, the training module 730 may also be for performing the operation S230 described above, which will not be explained here.

[0092] According to embodiments of the present disclosure, the device 700 may further include a second splitting module for splitting presentation information for a target object into at least one presentation word to obtain a presentation word sequence. The sequence acquisition module 720 specifically inputs the descriptive word sequence and the presentation word sequence into a dialogue generation model to obtain a probability vector sequence.

[0093] According to the embodiments of this disclosure, the probability vector sequence provides a second recommendation information for the target object. Show The device 700 may further include a probability identification module and a first loss identification module. The probability identification module is for identifying the predicted probability that the second recommendation information contains the presented information, based on a probability vector sequence. The first loss identification module is for identifying a first loss value of the dialogue generation model based on the predicted probability. The training module 730 is further used to train the dialogue generation model based on the first loss value.

[0094] According to embodiments of this disclosure, the probability identification module may include a candidate identification submodule and a probability identification submodule. The first identification submodule is for identifying the probability that the second recommendation information contains each suggested word, based on the probability value for each suggested word in the suggested word sequence from the probability vector sequence. The second identification submodule is for identifying the predicted probability that the second recommendation information contains the suggested information, based on at least one probability that the second recommendation information contains at least one suggested word.

[0095] According to embodiments of this disclosure, a probability vector sequence indicates a second recommendation for a target. The device 700 may further include a second loss identification module for identifying a second loss value of the dialogue generation model based on the relationship between the second recommendation information and the descriptive information. The training module 730 may further train the dialogue generation model based on the second loss value.

[0096] According to embodiments of this disclosure, the dialogue generation model may include a pre-trained dialogue generation model with discrete latent variables. The sequence acquisition module 720 specifically inputs random identifier information and a descriptive word sequence into the dialogue generation model to acquire association predictive values ​​and probability vector sequences corresponding to the random identifier information. Here, the association predictive values ​​indicate the association relationship between the second recommendation information and the descriptive information.

[0097] According to embodiments of this disclosure, the device 700 may further include an information identification module, a first vector identification module, and a third loss identification module. The information identification module is for identifying second recommendation information for a target object based on a probability vector sequence. The first vector identification module, in response to the presence of duplicate words in the second recommendation information, identifies a probability vector corresponding to the duplicate words in the probability vector sequence based on the positional information of the duplicate words in the second recommendation information, and sets it as the target probability vector. The third loss identification module is for identifying a third loss value for the dialogue generation model based on the target probability vector and the duplicate words. Here, the training module 730 is for further training the dialogue generation model based on the third loss value.

[0098] According to embodiments of the present disclosure, the device 700 may further include a third splitting module, a second vector identification module, and a fourth loss identification module. The third splitting module is for splitting the first recommendation information into at least one recommendation word to obtain a recommendation word sequence. The second vector identification module, in response to the presence of duplicate words in the recommendation word sequence, identifies a probability vector corresponding to the duplicate word in the probability vector sequence based on the positional information of the duplicate word in the recommendation word sequence, and sets it as the target probability vector. The fourth loss identification module is for identifying a third loss value of the dialogue generation model based on the target probability vector and the duplicate word. Here, the training module 730 is further used to train the dialogue generation model based on the third loss value.

[0099] According to the embodiments of this disclosure, the dialogue generation model includes a pre-trained dialogue generation model with discrete latent variables.

[0100] Figure 8 is a block diagram of the configuration of an information generation device according to an embodiment of the present disclosure.

[0101] As shown in Figure 8, the information generation device 800 of this embodiment may include a first division module 810, a sequence acquisition module 820, and an information identification module 830.

[0102] The first splitting module 810 is for splitting the descriptive information of the object to be recommended into at least one descriptive word to obtain a descriptive word sequence. In one embodiment, the first splitting module 810 may also be for performing the operation S610 described above, which will not be explained here.

[0103] The sequence acquisition module 820 inputs a descriptive word sequence into an information generation model to acquire a probability vector sequence for the object to be recommended. Here, each probability vector in the probability vector sequence contains probability values ​​for a plurality of predetermined words, and the information generation model may be obtained by training using the information generation model training device provided in this disclosure. In one embodiment, the sequence acquisition module 820 may be for performing the operation S620 described above, and this will not be explained here.

[0104] The information identification module 830 is used to identify recommendation information for targets to be recommended based on a probability vector sequence. In one embodiment, the information identification module 830 may also be used to perform the operation S630 described above, and this will not be explained here.

[0105] According to embodiments of the present disclosure, the device 800 may further include a second splitting module for splitting presentation information for a subject to be recommended into at least one presentation word to obtain a presentation word sequence. The sequence acquisition module 820 specifically inputs the descriptive word sequence and the presentation word sequence into a dialogue generation model to obtain a probability vector sequence.

[0106] According to embodiments of this disclosure, the dialogue generation model may include a pre-trained dialogue generation model with discrete latent variables. The sequence acquisition module 820 specifically inputs random identifier information and descriptive word sequences into the dialogue generation model to acquire association values ​​and probability vector sequences corresponding to the random identifier information. Here, the association values ​​indicate the relationship between recommendation information and descriptive information.

[0107] It should be explained that in the proposed technical solutions of this disclosure, the collection, storage, use, processing, transmission, provision, disclosure, and application of such user personal information all comply with the provisions of the relevant laws and regulations, employ necessary security measures, and do not violate public order and morals. In the proposed technical solutions of this disclosure, permission or consent from the user is obtained before acquiring or collecting any user personal information.

[0108] According to embodiments of the present disclosure, the present disclosure further provides electronic devices, readable storage media, and computer programs.

[0109] Figure 9 schematically shows a block diagram of an exemplary electronic device 900 capable of implementing the training method for the information generation model and / or the information generation method of an embodiment of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other appropriate computers. The electronic device may further represent various forms of mobile devices, such as personal digital assistants, mobile phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are illustrative and not intended to limit the implementation of the present disclosure as described herein and / or requested.

[0110] As shown in Figure 9, the electronic device 900 includes a computing means 901 that can perform various appropriate operations and processes based on a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage means 908 into a random access memory (RAM) 903. The RAM 903 may further store various programs and data necessary for the operation of the electronic device 900. The computing means 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is further connected to the bus 904.

[0111] Multiple components in the electronic device 900 are connected to an I / O interface 905 and include, for example, input means 906 such as a keyboard and mouse; output means 907 such as various types of displays and speakers; storage means 908 such as magnetic disks and optical disks; and communication means 909 such as a network card, modem, and wireless communication transceiver. The communication means 909 enables the electronic device 900 to exchange information / data with other devices via computer networks such as the Internet and / or various telecommunication networks.

[0112] The computing means 901 may be various general-purpose and / or dedicated processing modules having processing and computing capabilities. Some examples of the computing means 901 include, but are not limited to, a central processing unit (CPU), an image processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, computing means for executing various device learning model algorithms, a digital signal processor (DSP), any appropriate processor, controller, or microcontroller. The computing means 901 performs each of the above methods and processes, for example, an information generation model training method and / or information generation method. For example, in some embodiments, the information generation model training method and / or information generation method are implemented as a tangible computer software program contained in a device-readable medium, for example, a storage means 908. In some embodiments, part or all of the computer program is loaded and / or installed into the electronic device 900 via ROM 902 and / or communication means 909. When the computer program is loaded into RAM 903 and executed by the computing means 901, one or more steps of the above information generation model training method and / or information generation method can be performed. Alternatively, in other embodiments, the computing means 901 may be configured to perform a method for training an information generation model and / or an information generation method by any other suitable method (e.g., a firmware-based method).

[0113] Various embodiments of the systems and technologies described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), dedicated integrated circuits (ASICs), dedicated standard products (ASSPs), system-on-chip systems (SOCs), load-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include being executed by one or more computer programs, which may be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, which may receive data and instructions from a storage system, at least one input device, and at least one output device, and transfer data and instructions to the storage system, the at least one input device, and the at least one output device.

[0114] Program code for carrying out the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, a dedicated computer, or other programmable data processing device, so that when the program code is executed by the processor or controller, the functions and operations defined in the flowchart and / or block diagrams are performed. The program code may be executed entirely on a device, partially on a device, partially on a device and partially on a remote device as a standalone software pack, or entirely on a remote device or server.

[0115] In the context of this disclosure, a device-readable medium may be a tangible medium containing or storing a program used in or in combination with an instruction execution system, apparatus, or device. A device-readable medium may be a device-readable signal medium or a device-readable storage medium. A device-readable medium includes, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any appropriate combination thereof. More specific examples of device-readable storage media include one or more line electrical connections, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any appropriate combination thereof.

[0116] To provide user interaction, a computer may implement the systems and techniques described herein, which may have a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) and a keyboard and pointing device (e.g., a mouse or trackball), and the user may provide input to the computer using the keyboard and pointing device. Other types of devices may also be for providing user interaction, for example, the feedback provided to the user may be any form of sensor feedback (e.g., visual feedback, auditory feedback, or haptic feedback), and input from the user may be received in any form (including voice input, speech input, or haptic input).

[0117] The systems and technologies described herein may be implemented in a computing system including background components (e.g., a data server), a computing system including middleware components (e.g., an application server), or a computing system including front-end components (e.g., a user computer having a graphical user interface or a network browser, through which the user can interact with embodiments of the systems and technologies described herein), or in a computing system including any combination of such background components, middleware components, or front-end components. Components of the system may be interconnected via digital data communication (e.g., a communication network) of any form or medium. Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0118] A computer system may include a client and a server. The client and server are generally geographically separated and interact typically via a communication network. The client-server relationship is formed by computer programs running on corresponding computers and having a client-server relationship with each other. Here, the server may be a cloud server, also called a cloud computing server or cloud host, and is a host product within a cloud computing service system, thereby resolving the drawbacks of traditional physical hosts and VPS services ("Virtual Private Server," or "VPS"), such as high management difficulty and low service scalability. The server may be a server in a distributed system, or a server integrated with a blockchain.

[0119] It should be understood that steps can be rearranged, added, or deleted using the various forms of flows described above. For example, each step described in this disclosure may be performed in parallel, sequentially, or in a different order, as long as the disclosed technical proposal achieves the desired result, this specification is not limited thereto.

[0120] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, subcombinations, and substitutions can be made depending on design requirements and other factors. Any modifications, equivalent substitutions, improvements, etc., made within the scope of the intent and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for training information generation models, The method involves dividing the descriptive information for a target object in an information pair into at least one descriptive word to obtain a descriptive word sequence, wherein the information pair consists of descriptive information and recommendation information for the target object, and the recommendation information includes first recommendation information. The process involves inputting the descriptive word sequence into a dialogue generation model to obtain a probability vector sequence for the target object, wherein each probability vector in the probability vector sequence includes probability values ​​for a plurality of predetermined words. The dialogue generation model is trained by maximizing the probability of generating the first recommendation information based on the probability vector sequence and the first recommendation information, and the dialogue generation model that reaches a predetermined convergence condition is adopted as the information generation model. It includes, The aforementioned probability vector sequence indicates a second recommendation for the target object, Training the dialogue generation model based on the aforementioned probability vector sequence and the first recommendation information is: Based on the aforementioned probability vector sequence, the second recommendation information identifies the predicted probability that the presented information includes, Based on the aforementioned predicted probability, the first loss value of the dialogue generation model is identified, Based on the semantic similarity or thematic similarity, which is the relationship between the second recommendation information and the descriptive information, a second loss value of the dialogue generation model is identified. The dialogue generation model is trained based on the first loss value, the second loss value, and the first recommendation information. It includes, Both the presented information and the descriptive information are provided by the manufacturer or seller of the target product. Based on the aforementioned probability vector sequence, the second recommendation information identifies the predicted probability that the presented information includes the following: For each probability vector, a predetermined word corresponding to the maximum probability value in each of the said probability vectors is identified, and one predicted word is obtained. The obtained plurality of predicted words are sorted based on their position in the probability vector sequence of the probability vector, and the plurality of predicted words are combined to constitute the second recommendation information. This includes searching for the second recommendation information using each word in the presented word sequence, determining whether the second recommendation information contains each word, and defining the proportion of the second recommendation information that contains the words in the presented word sequence as the predicted probability that the second recommendation information contains the presented information. Training methods for information generation models.

2. The method further includes dividing the information presented to the target object into at least one presented word to obtain a sequence of presented words, Inputting the descriptive word sequence into a dialogue generation model to obtain a probability vector sequence for the target object includes inputting the descriptive word sequence and the presented word sequence into the dialogue generation model to obtain a probability vector sequence. A method for training an information generation model according to claim 1.

3. Based on the aforementioned probability vector sequence, the second recommendation information identifies a predicted probability that includes the presented information. Based on the probability values ​​for each suggested word in the suggested word sequence among the probability vector sequence, the probability that the second recommendation information includes each suggested word is identified. Based on the probability that the second recommendation information contains the at least one suggested word, the predicted probability that the second recommendation information contains the suggested information is identified. including A method for training an information generation model according to claim 2.

4. The dialogue generation model includes a pre-trained dialogue generation model with discrete latent variables, Inputting the aforementioned descriptive word sequence into the dialogue generation model to obtain a probability vector sequence for the target object is, This includes inputting random identification information and the descriptive word sequence into the dialogue generation model to obtain associated predicted values ​​and probability vector sequences corresponding to the random identification information, The aforementioned related predicted values ​​indicate the relationship between the second recommendation information and the descriptive information. A method for training an information generation model according to claim 1.

5. In the process of identifying the second recommendation information, in response to whether or not there are duplicate words among the multiple predicted words obtained, the probability vector corresponding to the duplicate word in the probability vector sequence is identified based on the positional information of the duplicate word in the second recommendation information, and this probability vector is set as the target probability vector. Based on the target probability vector and duplicate words, a third loss value of the dialogue generation model is identified. It includes, However, the third loss value is determined based on the maximum probability value contained in any one of the target probability vectors, which is the probability value for the duplicate word contained in any one of the target probability vectors. The third loss value has a positive correlation with the probability value for the duplicate word in the target probability vector. Training the dialogue generation model based on the aforementioned probability vector sequence and the first recommendation information is: The further includes training the dialogue generation model based on the first loss value, the second loss value, and the third loss value. A method for training an information generation model according to claim 1.

6. The first recommendation information is divided into at least one recommendation word to obtain a recommendation word sequence, In response to the presence of duplicate words in the aforementioned recommended word sequence, the probability vector corresponding to the duplicate word in the probability vector sequence is identified based on the positional information of the duplicate word in the aforementioned recommended word sequence, and this is set as the target probability vector. Based on the target probability vector and duplicate words, a third loss value of the dialogue generation model is identified. It includes, Training the dialogue generation model based on the aforementioned probability vector sequence and the first recommendation information is: The further includes training the dialogue generation model based on the first loss value, the second loss value, and the third loss value. A method for training an information generation model according to claim 1.

7. In information generation methods, The descriptive information of the subject to be recommended is divided into at least one descriptive word, and a sequence of descriptive words is obtained. The process involves inputting the aforementioned descriptive word sequence into an information generation model to obtain a probability vector sequence for the object to be recommended, wherein each probability vector in the probability vector sequence contains probability values ​​for a plurality of predetermined words. A dialogue generation model is trained, and the dialogue generation model that reaches a predetermined convergence condition is used as an information generation model. The probability vector sequence indicates recommendation information for the object to be recommended, and the recommendation information for the object to be recommended is identified by identifying the predicted probability that the recommendation information includes the presented information. Includes, The information generation model is trained using the information generation model training method described in claim 1. Information generation method.

8. The method further includes dividing the information presented for the subject to be recommended into at least one present word to obtain a present word sequence, Inputting the aforementioned descriptive word sequence into an information generation model to obtain a probability vector sequence for the object to be recommended is, This includes inputting the descriptive word sequence and the presented word sequence into the dialogue generation model to obtain the probability vector sequence. The information generation method according to claim 7.

9. The dialogue generation model includes a pre-trained dialogue generation model with discrete latent variables, Inputting the aforementioned descriptive word sequence into an information generation model to obtain a probability vector sequence for the object to be recommended is, This includes adding one identification bit to the input of the dialogue generation model, inputting the random identification information and the descriptive word sequence to which the identification bit is adopted into the dialogue generation model, and obtaining the associated value and the probability vector sequence corresponding to the random identification information, The aforementioned relationship value indicates the relationship of semantic similarity or thematic similarity between the recommendation information and the descriptive information. The information generation method according to claim 7.

10. A training device for information generation models, A first splitting module for obtaining a sequence of descriptive words by splitting descriptive information for a target object in an information pair into at least one descriptive word, wherein the information pair consists of descriptive information and recommendation information for the target object, and the recommendation information includes a first splitting module containing first recommendation information, A sequence acquisition module for inputting the aforementioned descriptive word sequence into a dialogue generation model to obtain a probability vector sequence for the target object, wherein each probability vector in the probability vector sequence includes probability values ​​for a plurality of predetermined words, and the probability vector sequence indicates second recommendation information for the target object. A probability identification module for identifying the predicted probability that the second recommendation information includes the presented information, based on the probability vector sequence, A first loss identification module for identifying a first loss value of the dialogue generation model based on the predicted probability, A second loss identification module for identifying a second loss value of the dialogue generation model based on semantic similarity or thematic similarity, which is the relationship between the second recommendation information and the descriptive information, A training module that trains the dialogue generation model by maximizing the probability of generating the first recommendation information based on the first loss value, the second loss value, and the first recommendation information, and uses the dialogue generation model that has reached a predetermined convergence condition as the information generation model, It includes, Both the presented information and the descriptive information are provided by the manufacturer or seller of the target product. The aforementioned probability determination module is For each probability vector, a predetermined word corresponding to the maximum probability value in each of the probability vectors is identified, and one predicted word is obtained. The obtained plurality of predicted words are sorted based on their position in the probability vector sequence of the probability vector, and the plurality of predicted words are combined to form the second recommendation information. The second recommendation information is searched using each word in the presented word sequence, it is determined whether the second recommendation information contains each word, and the proportion of the second recommendation information containing the words in the presented word sequence is defined as the predicted probability that the second recommendation information contains the presented information. A training device for information generation models.

11. The system further includes a second splitting module for splitting the information presented to the target object into at least one presented word to obtain a presented word sequence, The sequence acquisition module is used to input the descriptive word sequence and the presented word sequence into the dialogue generation model to acquire a probability vector sequence. A training device for an information generation model according to claim 10.

12. The aforementioned probability determination module is A first identification submodule for identifying the probability that the second recommendation information contains each of the suggested words, based on the probability values ​​for each suggested word in the suggested word sequence among the probability vector sequence, A second identifying submodule for identifying a predicted probability that the second recommendation information includes the suggested information, based on at least one probability that the second recommendation information includes the suggested word, including A training device for an information generation model according to claim 11.

13. The dialogue generation model includes a pre-trained dialogue generation model with discrete latent variables, The sequence acquisition module described above is Random identification information and the descriptive word sequence are input to the dialogue generation model to obtain associated predicted values ​​and probability vector sequences corresponding to the random identification information. The aforementioned related predicted values ​​indicate the relationship between the second recommendation information and the descriptive information. A training device for an information generation model according to claim 10.

14. A first vector identification module, which, in response to whether or not there are duplicate words among the multiple predicted words obtained in the process of identifying the second recommendation information, identifies a probability vector corresponding to the duplicate word in the probability vector sequence based on the positional information of the duplicate word in the second recommendation information, and sets this probability vector as the target probability vector, A third loss identification module for identifying a third loss value of the dialogue generation model based on the target probability vector and duplicate words, It further includes, However, the third loss value is determined based on the maximum probability value contained in any one of the target probability vectors, which is the probability value for the duplicate word contained in any one of the target probability vectors. The third loss value has a positive correlation with the probability value for the duplicate word in the target probability vector, The training module is further used to train the dialogue generation model based on the first loss value, the second loss value, and the third loss value. A training device for an information generation model according to claim 10.

15. A third splitting module for splitting the first recommendation information into at least one recommendation word to obtain a recommendation word sequence, A second vector identification module, which, in response to the presence of duplicate words in the recommended word sequence, identifies a probability vector corresponding to the duplicate word in the probability vector sequence based on the positional information of the duplicate word in the recommended word sequence and sets it as the target probability vector, A fourth loss identification module for identifying a third loss value of the dialogue generation model based on the target probability vector and duplicate words, Includes, The training module is further used to train the dialogue generation model based on the first loss value, the second loss value, and the third loss value. A training device for an information generation model according to claim 10.

16. An information generating device, A first splitting module for obtaining a sequence of descriptive words by splitting the descriptive information of the target to be recommended into at least one descriptive word, A sequence acquisition module for inputting the descriptive word sequence into an information generation model to obtain a probability vector sequence for the object to be recommended, wherein each probability vector in the probability vector sequence includes a sequence acquisition module containing probability values ​​for a plurality of predetermined words. A dialogue generation model is trained, and the dialogue generation model that reaches a predetermined convergence condition is used as an information generation model. The probability vector sequence indicates recommendation information for the object to be recommended, and the information identification module identifies the recommendation information for the object to be recommended by identifying the predicted probability that the recommendation information includes the presented information. Includes, The information generation model is trained using the information generation model training device described in any one of claims 10 to 15. Information generation device.

17. The system further includes a second splitting module for splitting the presentation information for the subject to be recommended into at least one presentation word to obtain a presentation word sequence, The sequence acquisition module is used to input the descriptive word sequence and the presented word sequence into the dialogue generation model to acquire the probability vector sequence. The information generating device according to claim 16.

18. The dialogue generation model includes a pre-trained dialogue generation model with discrete latent variables, The sequence acquisition module described above is This is used to add one identification bit to the input of the dialogue generation model, input the random identification information and the descriptive word sequence to the dialogue generation model, and obtain the associated value and the probability vector sequence corresponding to the random identification information. The aforementioned relationship value indicates the relationship of semantic similarity or thematic similarity between the recommendation information and the descriptive information. The information generating device according to claim 16.

19. It is an electronic device, At least one processor, A memory that is communicated with at least one of the aforementioned processors, Includes, The memory stores commands to be executed by at least one processor. An electronic device in which the command is executed on the at least one processor, thereby enabling the at least one processor to perform the training method for the information generation model described in claim 1.

20. It is an electronic device, At least one processor, A memory that is communicated with at least one of the aforementioned processors, Includes, The memory stores commands to be executed by at least one processor. An electronic device in which the command is executed on the at least one processor, thereby enabling the at least one processor to perform the information generation method according to claim 7.

21. A non-volatile, computer-readable storage medium storing computer commands for causing a computer to perform the training method for the information generation model described in claim 1.

22. A non-volatile computer-readable storage medium storing computer commands for causing a computer to perform the information generation method described in claim 7.

23. It is a computer program, A computer program that, when executed on a processor, realizes the training method for the information generation model described in claim 1.

24. It is a computer program, A computer program that implements the information generation method described in claim 7, when the computer program is executed on a processor.

Images