Text clustering method based on comparative learning and integrated with dynamic adjustment mechanism

Abstract

The invention discloses a text clustering method integrated with a dynamic adjustment mechanism based on comparative learning, and the method is characterized in that a group of enhanced texts is obtained based on a context enhancement method, the feature representation of the enhanced texts is obtained through a pre-training model, and the initial cluster center of a semantic cluster is obtained through a K-Means clustering method; the confidence coefficient of text clustering distribution is improved, and dynamic screening is carried out; and finally, obtaining a model total loss function, and continuously and dynamically adjusting through a dynamic adjustment function, so that the model training weight is smoothly transited to a clustering task from comparative learning. According to the invention, the problem of inconsistency between comparative learning and clustering targets is alleviated; smooth transition from comparative learning to clustering is realized; the negative examples are screened by allocating pseudo labels to the data with high confidence coefficient of the cluster allocation probability, so that the problem that the same cluster data are mutual negative examples is solved, and the quality of the negative examples is effectively improved; the data representation obtained by contrast learning is more friendly to clustering; the short text clustering method is superior to the existing short text clustering method in most data sets.

Description

technical field [0001] The invention belongs to the technical field of natural language processing, and in particular relates to a text clustering method based on comparative learning and integrating a dynamic adjustment mechanism. Background technique [0002] Text clustering is one of the unsupervised data processing methods. The purpose of this method is to divide the texts into different clusters according to the similarity without labels, so that the semantics of the texts in the clusters are as similar as possible, and the semantics of the texts outside the clusters are as similar as possible. different. In recent years, deep learning has attracted extensive attention, and a considerable part of clustering research work focuses on the combination of clustering and deep learning, using the powerful representation ability of deep learning to extract semantic features in text, and then clustering for better clustering effect. Yang et al. (2017) combined autoencoders and...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the above-mentioned technical problems, the present invention designs a text clustering method based on contrastive learning and incorporating a dynamic adjustment mechanism, in order to solve the low clustering semantic confidence caused by the inconsistency between auxiliary tasks and clustering task goals during deep text clustering problem, the present invention introduces the loss weight dynamic adjustment method and the contrastive learning negative example screening method into the text clustering model

Images