Cross-domain text sentiment classification method based on parameter migration and attention sharing mechanism

Abstract

The invention provideds a method and a system based on a parameter migration and attention sharing mechanism for a cross-domain text sentiment classification task. In particular, the present architecture includes a source domain network and a target domain network. Firstly, a hierarchical attention network with a pre-training language model is constructed on training data, and the pre-training language model comprises a global vector for word representation and a bidirectional encoder language model. Secondly, in model migration, a word and sentence level parameter migration mechanism is introduced, and network parameters are migrated from a source domain network to a target domain network by adopting a parameter migration and fine tuning technology. Finally, emotional attention can serve as a bridge for connecting emotional transmission of different fields, a word and sentence level attention mechanism is introduced, and cross-field emotional attention is shared from the two levels. Experiments show that the method provided by the invention achieves an optimal result on an Amazon cross-domain sentiment classification data set.

Description

technical field [0001] The invention relates to the field of natural language processing text sentiment analysis, and proposes a cross-domain text sentiment classification method based on a parameter migration and attention sharing mechanism. Background technique [0002] Traditional text sentiment classification methods assume that the fields used for training and testing are independent and identically distributed. However, under practical conditions, there are distributional differences between different domains. Cross-domain text sentiment classification uses source domain data resources to achieve sentiment classification tasks in the target domain. In order to effectively solve the problem of insufficient data labeling in specific domains, cross-domain sentiment classification extends the application of transfer learning in text-based social media, which can improve the classification effect of text sentiment classification tasks with insufficient data sources. Furth...

AI Technical Summary

Problems solved by technology

Existing methods mainly address two problems: (1) determine which parameters can be shared in the model? (2) How to share model parameters? That is, which method is used to realize the migration of model parameters

Images