An Online Learning Sorting Method Based on Listwise Algorithm

Abstract

The invention relates to the technical field of information retrieval, in particular to an online learning sorting method based on a listwise algorithm. First, using the idea of ​​the listwise algorithm, the sorted list of candidate records corresponding to the query keyword is used as a training example; then, using the training process of the online learning sorting algorithm, when the training example is applied, the training example arrives at the neural network in order and is scanned only once ; Finally, an accurate ranking model is obtained by minimizing the loss value. The present invention applies the online learning algorithm to the listwise algorithm, can use the data obtained online to update the existing sorting model, realize online processing and improve the effectiveness of the algorithm; ensure the implementation of the online-listwise algorithm in the field of information retrieval sex and performance. The ranking model is updated through an adaptive learning rate to make the ranking model more accurate and converge as soon as possible.

Description

technical field [0001] The invention relates to the technical field of information retrieval, in particular to an online learning and sorting method based on a listwise algorithm, which sorts candidate records corresponding to a current query keyword. Background technique [0002] In the field of information retrieval, given a query keyword, the search engine will recall a series of related candidate records, then sort the recalled candidate records, and finally output the first n candidate records. Therefore, how to effectively rank candidate records is a core issue of information retrieval. With the rapid development of machine learning algorithms, using machine learning algorithms to obtain accurate ranking models has been widely used. So far, the existing learning ranking algorithms are mainly divided into three categories: pointwise, pairwise, and listwise. [0003] The Pointwise algorithm uses machine learning algorithms to predict the sorting value of a single candi...

AI Technical Summary

Problems solved by technology

[0006] The defect of this prior art: the learning sorting algorithm proposed in this document, on the one hand, because the algorithm does not fully consider the internal dependencies of the candidate records corresponding to the query keywords in the training process, so when different query keywords correspond to When there are different numbers of candidate records, the overall loss value will be determined by the query keywords with a large amount of candidate records

[0008] The defect of this prior art: the learning sorting algorithm proposed in this document, on the one hand, needs to convert the candidate record corresponding to the query keyword into a candidate record pair when obtaining the training instance, so some may be lost in the conversion process. More fine-grained feature information, and if there is an error in the labeling of a candidate record pair during the conversion process, it may cause a chain reaction

[0010] The defect of this prior art: Although the learning sorting algorithm proposed in this document solves the problem of the location information of the candidate records corresponding to the query keywords, when calculating the loss value, it needs to sort the list based on the candidate records corresponding to the query keywords. calculation, and cannot update the ranking model online in real time

[0012] The defect of the existing technology: Although the Online-pairwise algorithm improves the effectiveness of the algorithm, there is still a large performance gap compared with the listwise algorithm in terms of performance

[0013] Among the many learning ranking algorithms in the field of information retrieval, the data obtained online cannot be used to update the ranking model, so as the amount of data used for information retrieval increases, the problem of poor effectiveness of learning ranking algorithms becomes more and more obvious

At the same time, the industry has not yet proposed an online learning sorting algorithm that uses the sorted list of candidate records corresponding to the query keyword as a training example

Images