Method and device for training unbiased peer-to-peer double-end learning sorting model, and medium

Abstract

The invention relates to an unbiased peer-to-peer double-end learning sorting model training method and device and a medium, and the method comprises the steps: constructing an unbiased peer-to-peer double-end learning sorting model which comprises a first relevancy error estimation model and a second relevancy error estimation model, and the first relevancy error estimation model and the second relevancy error estimation model are at least partially equal; and training and optimizing the unbiased peer-to-peer double-end learning sorting model by adopting an unbiased peer-to-peer double-end cross training mode based on a biased data set and by taking relevancy error reduction as a target. According to the unbiased peer-to-peer double-end learning sorting model obtained by the invention, various deviations caused by positions in a data set can be remarkably overcome, the matching degree of the query return document is improved, and the unbiased peer-to-peer double-end learning sorting model is high in practicability and suitable for various scenes such as internet search and service recommendation.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a training method, device, and medium for an unbiased peer-to-peer dual-end learning ranking model. Background Art

[0002] Unbiased learning to rank is a classic problem in the field of information retrieval and a core task in business scenarios such as Internet search and recommendation. Compared with traditional learning to rank methods, unbiased learning to rank can improve the accuracy of ranking to a certain extent and has been widely studied.

[0003] However, current unbiased learning-to-rank methods still suffer from widespread problems of various biases caused by the position of the service dataset. Summary of the invention

[0004] The present invention provides a training method, device, and medium for an unbiased peer-to-peer dual-end learning ranking model to overcome or at least partially overcome the deficiencies of the prior art.

[0005] In a first aspect, the present application provides a training method for an unbiased peer-to-peer dual-end learning ranking model, the method comprising:

[0006] Constructing an unbiased peer-to-peer dual-end learning ranking model, the unbiased peer-to-peer dual-end learning ranking model comprising: a first relevance error prediction model and a second relevance error prediction model that are at least partially equal;

[0007] Based on the biased data set and aiming at reducing the correlation error, an unbiased peer-to-peer two-end cross-training method is adopted to train and optimize the unbiased peer-to-peer two-end learning to sort model.

[0008] In a second aspect, the present application also provides a training device for an unbiased peer-to-peer dual-end learning ranking model, the device comprising:

[0009] A construction unit is used to construct an unbiased peer-to-peer dual-end learning to sort model, wherein the unbiased peer-to-peer dual-end learning to sort model includes: a first relevance error prediction model and a second relevance error prediction model that are at least partially equal;

[0010] The training unit is used to train and optimize the unbiased peer-to-peer dual-end learning ranking model based on a biased data set with the goal of reducing the correlation error by adopting an unbiased peer-to-peer dual-end cross-training method.

[0011] In a third aspect, the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of any of the above-mentioned training methods for the unbiased peer-to-peer two-end learning and sorting model.

[0012] The present application constructs an unbiased peer-to-peer two-end learning ranking model, which is a deep learning ranking model based on unbiased peer-to-peer two-end learning. The model includes two at least partially equivalent relevance error prediction models, which are respectively recorded as a first relevance error prediction model and a second relevance error prediction model. During training, based on a public biased data set and with the goal of reducing the relevance error, an unbiased peer-to-peer two-end cross-training method is adopted to train and optimize the model. The obtained unbiased peer-to-peer two-end learning ranking model can significantly overcome various deviations caused by position in the data set, improve the matching degree of query returned documents, has strong practicality, and is suitable for various scenarios such as Internet search and business recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0014] Figure 1 It is a flowchart of a training method for an unbiased peer-to-peer dual-end learning ranking model in one embodiment of the present invention;

[0015] Figure 2 A schematic diagram of the structure of an unbiased peer-to-peer dual-end learning ranking model according to an embodiment of the present application is shown;

[0016] Figure 3 It is a structural schematic diagram of a training device for an unbiased peer-to-peer dual-end learning to sort model in one embodiment of the present invention;

[0017] Figure 4 is a schematic diagram of the structure of a computer device in one embodiment of the present invention;

[0018] Figure 5 It is a schematic diagram of the structure of a computer device in another embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] In view of the problem of low sorting accuracy in the prior art due to various deviations caused by position in the data set, the present application proposes a training method for an unbiased peer-to-peer two-end learning sorting model, which mainly adopts two relatively equal models and an unbiased peer-to-peer two-end cross-learning approach to overcome the above problem.

[0021] Figure 1 A flow chart showing a training method for an unbiased peer-to-peer dual-end learning ranking model in an embodiment of the present application is shown, from Figure 1 It can be seen that this embodiment includes steps S110 to S120:

[0022] Step S110: constructing an unbiased peer-to-peer dual-end learning to rank model, wherein the unbiased peer-to-peer dual-end learning to rank model comprises: a first relevance error prediction model and a second relevance error prediction model that are at least partially equal.

[0023] Firstly, an unbiased peer-to-peer two-end learning ranking model is constructed. The model is a deep learning ranking model based on unbiased peer-to-peer two-end learning, which includes two at least partially equivalent correlation error estimation models.

[0024] Figure 2 The structural diagram of the unbiased peer-to-peer two-end learning ranking model according to an embodiment of the present application is shown. Figure 2 It can be seen that the unbiased equal dual-end learning to rank model 100 includes a first relevance error estimation model 110 and a second relevance error estimation model 120. The first relevance error estimation model 110 and the second relevance error estimation model 120 are at least equal, and the so-called "equal" means that the input, structure (such as model backbone, encoder, decoder, etc.), output, etc. of the model are consistent.

[0025] Preferably, in some embodiments of the present application, the first correlation error prediction model and the second correlation error prediction model are completely equivalent models in terms of input, model structure, and output. In other words, the two correlation error prediction models are completely consistent in all aspects and achieve "equivalence" in a strict sense.

[0026] Step S120: Based on the biased data set, with the goal of reducing the correlation error, an unbiased peer-to-peer dual-end cross-training method is adopted to train and optimize the unbiased peer-to-peer dual-end learning to rank model.

[0027] Then enter the training step. The training process in this application uses a biased data set. Specifically, any manually labeled data set in the prior art can be used. This data set is usually biased. A biased data set means that in the process of collecting or selecting data, some features or groups are over-represented, while other features or groups are underestimated or ignored. This deviation may cause the model to fail to correctly reflect the actual situation during training and testing, thereby affecting the accuracy and reliability of the prediction. Since the manual labeling is usually a certain group and has a strong subjective factor, this data set is usually biased, such as the expert data set in the public data set (A Large Scale Search Dataset for Unbiased Learning to Rank) released by Baidu.

[0028] In some embodiments of the present application, the biased data set has the following characteristics, which includes a training set and a validation set, the training set includes multiple training samples, each of the training samples includes a query and corresponding multiple documents, the query and each of the documents form a query-document pair, each of the query-document pairs has the same number of semantic features, and each of the query-document pairs has a relevance label value.

[0029] If the biased data set is denoted as set S, a training sample in set S is {Query_i, Document_1, Document_2,..., Document_k}, Query is the query, Document is the document; Query_i and Document_1, Document_2,..., Document_k form query-document pairs Query_i-Document_j (i=1,2,3,...; j=1,2,3,...k) respectively, and the semantic features f_k (k=1,2,3,...,M) of each query-document pair total M features, and each query-document pair has a relevance label value.

[0030] During training, overall, with the goal of reducing the correlation error, an unbiased peer-to-peer two-end cross-training method is adopted to train and optimize the unbiased peer-to-peer two-end learning to sort model, and the two peer correlation error prediction models of the unbiased peer-to-peer two-end learning to sort model are optimized and parameter adjusted in the direction of reducing their correlation error.

[0031] Specifically, based on the biased data set, with the goal of reducing the relevance error, the unbiased peer-to-peer dual-end learning ranking model is trained and optimized by using an unbiased peer-to-peer dual-end cross-training method, including: batch inputting multiple training samples in the biased data set into the unbiased peer-to-peer dual-end learning ranking model; for one training sample, respectively using the first relevance error prediction model and the second relevance error prediction model for reasoning to obtain a first relevance prediction value sequence and a second relevance prediction value sequence; respectively weighting the first relevance prediction value sequence and the second relevance prediction value sequence to obtain a first relevance prediction value weight sequence and a second relevance prediction value weight sequence; using the first relevance error prediction model, based on the second relevance prediction value sequence, multiple query sequences in the training sample are weighted. The relevance label values ​​of the document pairs are normalized to obtain a first relevance label value weight sequence; and the second relevance error prediction model is used to normalize the relevance label values ​​of multiple query-document pairs in the training sample based on the first relevance prediction value sequence to obtain a second relevance label value weight sequence; based on the first relevance prediction value sequence, the second relevance prediction value sequence, the first relevance prediction value weight sequence, the second relevance prediction value weight sequence, the first relevance label value weight sequence, and the second relevance label value weight sequence, the comprehensive error loss of the unbiased peer-to-peer dual-end learning to rank model is determined; and the gradient descent algorithm is used to optimize the comprehensive error loss based on the validation set to obtain a trained unbiased peer-to-peer dual-end learning to rank model.

[0032] Step 1: Input multiple training samples in the biased data set into the unbiased peer-to-peer dual-end learning ranking model in batches. Specifically, input multiple training samples into the first relevance error prediction model 110 and the second relevance error prediction model 120 in batches, respectively.

[0033] Step 2: For a training sample, the first correlation error prediction model and the second correlation error prediction model are used for inference to obtain a first correlation prediction value sequence and a second correlation prediction value sequence. The same process is performed for each training sample, and a training sample is used as an example for explanation.

[0034] Taking a training sample as an example, the first relevance error prediction model (denoted as end A) processes the semantic features of k query-document pairs corresponding to the query q in the input training sample, obtains the relevance prediction value of each query-document pair, and forms a relevance prediction value sequence, denoted as the first relevance prediction value, expressed as: Ar1, Ar2,..., Ark.

[0035] Similarly, taking a training sample as an example, the second relevance error prediction model (denoted as end B) processes the semantic features of k query-document pairs corresponding to the query q in the input training sample, obtains the relevance prediction value of each query-document pair, and forms a relevance prediction value sequence, denoted as the second relevance prediction value, expressed as: Br1, Br2, ..., Brk.

[0036] Step 3: further perform correlation prediction value weight calculation based on the correlation prediction value sequence to obtain a corresponding correlation prediction value weight sequence. Specifically, weighting is performed on the first correlation prediction value sequence and the second correlation prediction value sequence respectively to obtain a first correlation prediction value weight sequence and a second correlation prediction value weight sequence.

[0037] More specifically, the weighting processing is performed on the first relevance prediction value sequence and the second relevance prediction value sequence respectively to obtain a first relevance prediction value weight sequence and a second relevance prediction value weight sequence, including: normalizing and weighting the multiple first relevance prediction values ​​in the first relevance prediction value sequence to obtain the first relevance prediction value weight sequence, wherein each element in the first relevance prediction value weight sequence is a one-to-one corresponding first relevance prediction value weight value of each query-document pair; normalizing and weighting the multiple second relevance prediction values ​​in the second relevance prediction value sequence to obtain the second relevance prediction value weight sequence, wherein each element in the second relevance prediction value weight sequence is a one-to-one corresponding second relevance prediction value weight value of each query-document pair.

[0038] The first correlation prediction value sequence Ar1, Ar2, ..., Ark output by the first correlation error prediction model (end A) is subjected to softmax normalization calculation, and the calculation formula used is shown in formula (1):

[0039]

[0040] The calculation results are recorded as a sequence: As1, As2, ..., Ask.

[0041] The first correlation prediction value weight sequence is further obtained by weight calculation: Aw1 = As1 / As1, Aw2 = As1 / As2, ..., Awk = As1 / Ask. Similarly, the second correlation prediction value sequence Br1, Br2, ..., Brk output by the second correlation error estimation model (B end) is subjected to softmax normalization calculation, and the calculation formula used is still as shown in formula (1).

[0042] The calculation results are recorded as a sequence: Bs1, Bs2, ..., Bsk.

[0043] The second correlation prediction value weight sequence is further obtained through weight calculation: Bw1=Bs1 / Bs1, Bw2=Bs1 / Bs2, ..., Bwk=Bs1 / Bsk.

[0044] Step 4: Use a "cross" approach to weight the label values ​​of the training samples in the biased data set. Specifically, use the first relevance error prediction model to normalize the relevance label values ​​of multiple query-document pairs in the training sample based on the second relevance prediction value weight sequence to obtain a first relevance label value weight sequence; and use the second relevance error prediction model to normalize the relevance label values ​​of multiple query-document pairs in the training sample based on the first relevance prediction value weight sequence to obtain a second relevance label value weight sequence.

[0045] More specifically, the first relevance error prediction model is used to normalize the relevance label values ​​of multiple query-document pairs in the training sample based on the second relevance prediction value weight sequence to obtain a first relevance label value weight sequence, including: using a gradient descent calculation method, cyclically executing for one training sample, using the first relevance error prediction model, based on the second relevance prediction value weight sequence, each of the query-document pairs is re-weighted to obtain a re-weighted weight sequence; and normalizing the re-weighted weight sequence until a first relevance label value weight sequence that meets the convergence condition is obtained.

[0046] For a query q, the relevance label sequence in the biased data set corresponding to the query q is recorded as: T1, T2, ..., Tk. The first relevance error estimation model (end A) assigns weights to the sequences T1, T2, ..., Tk based on the second relevance prediction value weight sequence Bw1, Bw2, ..., Bwk calculated as the above results, to obtain a reweighted weight sequence: TBw1 = T1*Bw1, TBw2 = T2*Bw2, ..., TBwk = Tk*Bwk. The first relevance error estimation model (end A) further normalizes the reweighted weight sequence TBw1, TBw2, ..., TBwk to obtain the normalized weight of the weighted relevance, recorded as the first relevance label value weight sequence, which is expressed as follows:

[0047] TBNw1=TBw1 / (TBw1+TBw2+,...,+TBwk),TBNw2=TBw2 / (TBw1+TBw2+,...,+TBwk),...,TBNwk=TBwk / (TBw1+TBw2+,...,+TBwk); adopt gradient descent method for loop calculation. If TBw1, TBw2,...,TBwk are all 0, then TBNw1=1 / k, TBNw2=1 / k,...,TBNwk=1 / k. At this time, the first correlation label value weight sequence is the final result, that is, the loop is ended and no gradient calculation is performed.

[0048] Similarly, for a query q, the relevance label sequence in the biased data set corresponding to the query q is recorded as: T1, T2, ..., Tk. The second relevance error estimation model (side B) assigns weights to the sequences T1, T2, ..., Tk based on the above-mentioned calculation results of the first relevance prediction value weight sequence Aw1, Aw2,, ..., Awk, and obtains a re-weighted weight sequence: TAw1=T1*Aw1, TAw2=T2*Aw2, ..., TAwk=Tk*Awk. The second relevance error estimation model (side B) further normalizes the re-weighted weight sequence TAw1=T1*Aw1, TAw2=T2*Aw2, ..., TAwk=Tk*Awk to obtain The normalized weight of the relevance of the weight is recorded as the second relevance label value weight sequence, which is expressed as follows: TANw1=TAw1 / (TAw1+TAw2+,...,+TAwk), TANw2=TAw2 / (TAw1+TAw2+,...,+TAwk),...,TANwk=TAwk / (TAw1+TAw2+,...,+TAwk), and the gradient descent method is used for loop calculation. If TAw1, TAw2,...,TAwk are all 0, then TANw1=1 / k, TANw2=1 / k,...,TANwk=1 / k. At this time, the second relevance label value weight sequence is the final result, that is, the loop is ended and no gradient calculation is performed.

[0049] Step 5, calculation of loss function. The overall error loss function of the unbiased peer-to-peer dual-end learning to sort model is recorded as a comprehensive error loss, which includes two parts, namely the error loss of the first relevance error estimation model (end A) and the second relevance error estimation model (end B). Specifically, based on the first relevance prediction value sequence, the second relevance prediction value sequence, the first relevance label value weight sequence, and the second relevance label value weight sequence, the comprehensive error loss of the unbiased peer-to-peer dual-end learning to sort model is determined.

[0050] More specifically, the comprehensive error loss of the unbiased peer-to-peer dual-end learning and ranking model is determined based on the first relevance prediction value sequence, the second relevance prediction value sequence, the first relevance label value weight sequence, and the second relevance label value weight sequence, including: determining the first prediction value loss sequence of the first relevance error estimation model based on the first relevance prediction value sequence; and determining the second prediction value loss sequence of the second relevance error estimation model based on the second relevance prediction value sequence; multiplying the first prediction value loss sequence with the first relevance label value weight sequence to obtain a first sub-comprehensive error loss sequence of the first relevance error estimation model; multiplying the second prediction value loss sequence with the second relevance label value weight sequence to obtain a second sub-comprehensive error loss sequence of the second relevance error estimation model; and summing the first sub-comprehensive error loss sequence with the second sub-comprehensive error loss sequence to obtain the comprehensive error loss of the unbiased peer-to-peer dual-end learning and ranking model.

[0051] Based on the first correlation prediction value sequence Ar1, Ar2, ..., Ark of the aforementioned calculation results, the first correlation error prediction model (end A) performs log-softmax calculation on it, and the calculation formula is shown in formula (2):

[0052]

[0053] The calculation result is the first prediction value loss sequence ALSr1, ALSr2, ..., ALSrk.

[0054] The first relevance label value weight sequence TBNw1, TBNw2, ..., TBNwk and the first prediction value loss sequence ALSr1, ALSr2, ..., ALSrk are multiplied to obtain the comprehensive error loss of the first relevance error estimation model (end A), which is recorded as the first sub-comprehensive error loss sequence ATotalLoss, as shown in formula (3):

[0055] ATotalLoss=TBNw1*ALSr1+TBNw2*ALSr2+,...,+TBNwk*ALSrk formula (3).

[0056] Based on the second correlation prediction value sequence Br1, Br2, ..., Brk of the aforementioned calculation results, the second correlation error prediction model (B side) performs log-softmax calculation on it, and the calculation formula is formula (2). The calculation result is the second prediction value loss sequence BLSr1, BLSr2, ..., BLSrk.

[0057] The second relevance label value weight sequence TANw1,TANw2,...,TANwk and the second prediction value loss sequence BLSr1,BLSr2,...,BLSrk are multiplied to obtain the comprehensive error loss of the second relevance error estimation model (side B), which is recorded as the second sub-comprehensive error loss sequence BTotalLoss, as shown in formula (4):

[0058] BTotalLoss=TANw1*BLSr1+TANw2*BLSr2+,...,+TANwk*BLSrk formula (4).

[0059] Finally, the first sub-comprehensive error loss sequence ATotalLoss is summed with the second sub-comprehensive error loss sequence BTotalLoss to obtain the comprehensive error loss of the unbiased peer-to-peer two-end learning ranking model, as shown in formula (4):

[0060] TotalLoss=ATotalLoss+BTotalLoss formula (5).

[0061] Step six, cyclic convergence training. Specifically, the gradient descent algorithm is used to optimize the comprehensive error loss, and verification is performed based on the verification set to obtain a trained unbiased peer-to-peer two-end learning and sorting model. More specifically, for the comprehensive error loss value of the current training round, any verification sample in the verification set is used for verification. If the comprehensive error loss value obtained is the latest maximum value, the number of times convergence is satisfied is increased by one, until the number of times convergence is satisfied reaches a preset threshold, and the training is terminated to obtain a trained unbiased peer-to-peer two-end learning and sorting model.

[0062] The model is optimized using the stochastic gradient descent optimization algorithm. A comprehensive error loss TotalLoss is obtained after one training round. The error size is then verified in the validation set. If the error size in the validation set does not reach a new low, the number of times it does not reach a new low is increased by one until the set number of times is reached, and the training is terminated. Otherwise, steps one to five are repeated until the above requirements are met, and the training is terminated to obtain a trained unbiased peer-to-peer dual-end learning ranking model.

[0063] Depend on Figure 1It can be seen from the method shown that the present application constructs an unbiased peer-to-peer two-end learning ranking model, which is a deep learning ranking model based on unbiased peer-to-peer two-end learning. The model includes two at least partially equivalent relevance error prediction models, which are respectively recorded as the first relevance error prediction model and the second relevance error prediction model; during training, based on a public biased data set and with the goal of reducing the relevance error, an unbiased peer-to-peer two-end cross-training method is adopted to train and optimize it. The obtained unbiased peer-to-peer two-end learning ranking model can significantly overcome various deviations caused by position in the data set, improve the matching degree of query returned documents, has strong practicality, and is suitable for various scenarios such as Internet search and business recommendation.

[0064] Figure 3 The structural diagram of the training device of the unbiased peer-to-peer dual-end learning ranking model according to an embodiment of the present application is shown. Figure 3 It can be seen that the training device 300 of the unbiased peer-to-peer dual-end learning to sort model includes:

[0065] A construction unit 310 is used to construct an unbiased peer-to-peer dual-end learning to rank model, wherein the unbiased peer-to-peer dual-end learning to rank model includes: a first relevance error prediction model and a second relevance error prediction model that are at least partially equal;

[0066] The training unit 320 is used to train and optimize the unbiased peer-to-peer dual-end learning to rank model based on the biased data set and with the goal of reducing the correlation error by adopting an unbiased peer-to-peer dual-end cross-training method.

[0067] In some embodiments of the present application, in the above-mentioned device, the first correlation error prediction model and the second correlation error prediction model are completely equivalent models in terms of input, model structure, and output.

[0068] In some embodiments of the present application, in the above-mentioned device, the biased data set includes a training set and a validation set, the training set includes multiple training samples, each of the training samples includes a query and corresponding multiple documents, the query and each of the documents form a query-document pair, each of the query-document pairs has the same number of semantic features, and each of the query-document pairs has a relevance label value.

[0069] In some embodiments of the present application, in the above-mentioned device, the training unit 320 is used to batch input multiple training samples in the biased data set into the unbiased peer-to-peer two-end learning ranking model; for one training sample, the first relevance error prediction model and the second relevance error prediction model are used for reasoning respectively to obtain a first relevance prediction value sequence and a second relevance prediction value sequence; the first relevance prediction value sequence and the second relevance prediction value sequence are weighted respectively to obtain a first relevance prediction value weight sequence and a second relevance prediction value weight sequence; the first relevance error estimation model is used to calculate the relevance of multiple query-document pairs in the training sample based on the second relevance prediction value weight sequence. The label values ​​are normalized to obtain a first relevance label value weight sequence; and the second relevance error prediction model is used to normalize the relevance label values ​​of multiple query-document pairs in the training sample based on the first relevance prediction value weight sequence to obtain a second relevance label value weight sequence; based on the first relevance prediction value sequence, the second relevance prediction value sequence, the first relevance label value weight sequence, and the second relevance label value weight sequence, the comprehensive error loss of the unbiased peer-to-peer two-end learning to sort model is determined; the gradient descent algorithm is used to optimize the comprehensive error loss, and verification is performed based on the verification set to obtain a trained unbiased peer-to-peer two-end learning to sort model.

[0070] In some embodiments of the present application, in the above-mentioned device, the training unit 320 is used to normalize and weight multiple first relevance prediction values ​​in the first relevance prediction value sequence to obtain the first relevance prediction value weight sequence, and each element in the first relevance prediction value weight sequence is a one-to-one corresponding first relevance prediction value weight value of each query-document pair; and normalize and weight multiple second relevance prediction values ​​in the second relevance prediction value sequence to obtain the second relevance prediction value weight sequence, and each element in the second relevance prediction value weight sequence is a one-to-one corresponding second relevance prediction value weight value of each query-document pair.

[0071] In some embodiments of the present application, in the above-mentioned device, the training unit 320 is used to adopt a gradient descent calculation method to cyclically execute the steps of, for a training sample, using the first relevance error prediction model, based on the second relevance prediction value weight sequence, re-weighting each of the query-document pairs to obtain a re-weighted weight sequence; normalizing the re-weighted weight sequence until a first relevance label value weight sequence that meets the convergence condition is obtained; adopt a gradient descent calculation method to cyclically execute the steps of, for a training sample, using the second relevance error prediction model, based on the first relevance prediction value weight sequence, re-weighting each of the query-document pairs to obtain a re-weighted weight sequence; normalizing the re-weighted weight sequence until a second relevance label value weight sequence that meets the convergence condition is obtained.

[0072] In some embodiments of the present application, in the above-mentioned device, the training unit 320 is used to determine the first prediction value loss sequence of the first correlation error estimation model based on the first correlation prediction value sequence; and determine the second prediction value loss sequence of the second correlation error estimation model based on the second correlation prediction value sequence; multiply the first prediction value loss sequence with the first correlation label value weight sequence to obtain the first sub-comprehensive error loss sequence of the first correlation error estimation model; multiply the second prediction value loss sequence with the second correlation label value weight sequence to obtain the second sub-comprehensive error loss sequence of the second correlation error estimation model; and sum the first sub-comprehensive error loss sequence with the second sub-comprehensive error loss sequence to obtain the comprehensive error loss of the unbiased peer-to-peer dual-end learning ranking model.

[0073] In some embodiments of the present application, in the above-mentioned device, the training unit 320 is used to verify the comprehensive error loss value of the current training round using any verification sample in the verification set. If the comprehensive error loss value obtained is the latest maximum value, the number of times convergence is satisfied is increased by one until the number of times convergence reaches a preset threshold, and the training is terminated to obtain a trained unbiased peer-to-peer two-end learning sorting model.

[0074] It should be noted that the training device of the above-mentioned unbiased peer-to-peer two-end learning to sort model can implement the training method of the above-mentioned unbiased peer-to-peer two-end learning to sort model one-to-one, which will not be repeated here.

[0075] In one embodiment, a computer device is provided. The computer device may be a service line server. The internal structure diagram thereof may be as follows: Figure 4As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, the steps of a training method for an unbiased peer-to-peer dual-end learning sorting model are implemented.

[0076] In one embodiment, a computer device is provided. The computer device may be a client, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server through a network connection. When the computer program is executed by the processor, the steps of a training method for an unbiased peer-to-peer dual-end learning sorting model are implemented.

[0077] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps of a training method for an unbiased peer-to-peer dual-end learning to sort model are implemented.

[0078] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the following training method of an unbiased peer-to-peer two-end learning to sort model are implemented.

[0079] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can refer to the relevant descriptions of the business line service side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0080] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0081] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0082] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A training method for an unbiased peer-to-peer dual-end learning ranking model, characterized in that: The method comprises: Constructing an unbiased peer-to-peer dual-end learning ranking model, the unbiased peer-to-peer dual-end learning ranking model comprising: a first relevance error prediction model and a second relevance error prediction model that are at least partially equal; Based on the biased data set and aiming at reducing the correlation error, an unbiased peer-to-peer two-end cross-training method is adopted to train and optimize the unbiased peer-to-peer two-end learning to sort model.

2. The training method of the unbiased peer-to-peer dual-end learning ranking model according to claim 1 is characterized in that: The first correlation error prediction model and the second correlation error prediction model are completely equivalent models in terms of input, model structure, and output.

3. The training method of the unbiased peer-to-peer dual-end learning ranking model according to claim 1 is characterized in that: The biased data set includes a training set and a validation set, the training set includes multiple training samples, each of the training samples includes a query and corresponding multiple documents, the query and each of the documents form a query-document pair, each of the query-document pairs has the same number of semantic features, and each of the query-document pairs has a relevance label value.

4. The training method of the unbiased peer-to-peer dual-end learning ranking model according to claim 3 is characterized in that: Based on the biased data set, with the goal of reducing the correlation error, an unbiased peer-to-peer dual-end cross-training method is adopted to train and optimize the unbiased peer-to-peer dual-end learning ranking model, including: Batch inputting multiple training samples in the biased data set into the unbiased peer-to-peer dual-end learning ranking model; For a training sample, respectively using the first correlation error prediction model and the second correlation error prediction model to perform reasoning to obtain a first correlation prediction value sequence and a second correlation prediction value sequence; Performing weighting processing on the first relevance prediction value sequence and the second relevance prediction value sequence respectively to obtain a first relevance prediction value weight sequence and a second relevance prediction value weight sequence; Using the first relevance error prediction model, based on the second relevance prediction value weight sequence, normalize the relevance label values ​​of multiple query-document pairs in the training sample to obtain a first relevance label value weight sequence; and using the second relevance error prediction model, based on the first relevance prediction value weight sequence, normalize the relevance label values ​​of multiple query-document pairs in the training sample to obtain a second relevance label value weight sequence; Determining a comprehensive error loss of the unbiased peer-to-peer two-end learning to rank model based on the first relevance prediction value sequence, the second relevance prediction value sequence, the first relevance label value weight sequence, and the second relevance label value weight sequence; The comprehensive error loss is optimized by using a gradient descent algorithm, and is verified based on the verification set to obtain a trained unbiased peer-to-peer dual-end learning ranking model.

5. The training method of the unbiased peer-to-peer dual-end learning ranking model according to claim 4 is characterized in that: The step of respectively performing weighting processing on the first relevance prediction value sequence and the second relevance prediction value sequence to obtain a first relevance prediction value weight sequence and a second relevance prediction value weight sequence comprises: Normalizing and weighting a plurality of first relevance prediction values ​​in the first relevance prediction value sequence to obtain the first relevance prediction value weight sequence, wherein each element in the first relevance prediction value weight sequence is a first relevance prediction value weight value corresponding to each query-document pair one by one; The multiple second relevance prediction values ​​in the second relevance prediction value sequence are normalized and weighted to obtain the second relevance prediction value weight sequence, wherein each element in the second relevance prediction value weight sequence is a second relevance prediction value weight value corresponding one-to-one to each query-document pair.

6. The training method of the unbiased peer-to-peer dual-end learning ranking model according to claim 4 is characterized in that: The first relevance error prediction model is used to normalize the relevance label values ​​of multiple query-document pairs in the training sample based on the second relevance prediction value weight sequence to obtain a first relevance label value weight sequence, including: Adopting a gradient descent calculation method, cyclically executing the steps of re-weighting each query-document pair based on the second relevance prediction value weight sequence and the first relevance error estimation model for a training sample to obtain a re-weighted weight sequence; and normalizing the re-weighted weight sequence until a first relevance label value weight sequence that meets a convergence condition is obtained; The second relevance error prediction model is used to normalize the relevance label values ​​of multiple query-document pairs in the training sample based on the first relevance prediction value weight sequence to obtain a second relevance label value weight sequence, including: Adopting the gradient descent calculation method, the steps of cyclically executing, for a training sample, using the second relevance error prediction model, re-weighting each query-document pair based on the first relevance prediction value weight sequence to obtain a re-weighted weight sequence; and normalizing the re-weighted weight sequence until a second relevance label value weight sequence that meets the convergence condition is obtained.

7. The training method of the unbiased peer-to-peer dual-end learning ranking model according to claim 4 is characterized in that: The step of determining the comprehensive error loss of the unbiased peer-to-peer dual-end learning to rank model based on the first relevance prediction value sequence, the second relevance prediction value sequence, the first relevance label value weight sequence, and the second relevance label value weight sequence includes: Determining a first prediction value loss sequence of the first correlation error prediction model based on the first correlation prediction value sequence; and determining a second prediction value loss sequence of the second correlation error prediction model based on the second correlation prediction value sequence; Multiplying the first prediction value loss sequence by the first relevance label value weight sequence to obtain a first sub-comprehensive error loss sequence of the first relevance error estimation model; Multiplying the second prediction value loss sequence by the second relevance label value weight sequence to obtain a second sub-comprehensive error loss sequence of the second relevance error estimation model; The first sub-comprehensive error loss sequence is summed with the second sub-comprehensive error loss sequence to obtain the comprehensive error loss of the unbiased peer-to-peer two-end learning to sort model.

8. The training method of the unbiased peer-to-peer dual-end learning ranking model according to claim 4, characterized in that: The gradient descent algorithm is used to optimize the comprehensive error loss, and verification is performed based on the verification set to obtain a trained unbiased peer-to-peer two-end learning ranking model, including: For the comprehensive error loss value of the current training round, any verification sample in the verification set is used for verification. If the comprehensive error loss value obtained is the latest maximum value, the number of times convergence is satisfied is increased by one until the number of times convergence reaches the preset threshold, and the training is terminated to obtain a trained unbiased peer-to-peer two-end learning sorting model.

9. A training device for an unbiased peer-to-peer dual-end learning ranking model, characterized in that: The device comprises: A construction unit is used to construct an unbiased peer-to-peer dual-end learning to sort model, wherein the unbiased peer-to-peer dual-end learning to sort model includes: a first relevance error prediction model and a second relevance error prediction model that are at least partially equal; The training unit is used to train and optimize the unbiased peer-to-peer dual-end learning ranking model based on a biased data set with the goal of reducing the correlation error by adopting an unbiased peer-to-peer dual-end cross-training method.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the training method of the unbiased peer-to-peer dual-end learning to sort model as described in any one of claims 1 to 8 are implemented.

Images