Transfer Learning for Cross-Lingual SentimentrnClassification with Weakly Shared Deep Neural Networks

摘要

Cross-lingual sentiment classification aims to automaticallyrnpredict sentiment polarity (e.g., positive or negative) of datarnin a label-scarce target language by exploiting labeled datarnfrom a label-rich language. The fundamental challenge ofrncross-lingual learning stems from a lack of overlap betweenrnthe feature spaces of source language data and that of targetrnlanguage data. To address this challenge, previous studiesrnhave been performed to make use of the translated resourcesrnfor sentiment classification in the target language, and thernclassification performance is far from satisfactory becausernof the language gap between the source language and therntranslated target language.rnIn this paper, to address the above challenge, we presentrna novel deep neural network structure, called Weakly SharedrnDeep Neural Networks (WSDNNs), to transfer the crosslingualrninformation from a source language to a target language.rnTo share the sentiment labels between two languages,rnwe build multiple weakly shared layers of features. It allowsrnto represent both shared inter-language features andrnlanguage-specific ones, making this structure more flexiblernand powerful in capturing the feature representations of bilingualrnlanguages jointly. We conduct a set of experiments withrncross-lingual sentiment classification tasks on multilingualrnAmazon product reviews. The empirical results show thatrnour proposed approach significantly outperforms the stateof-rnthe-art methods for cross-lingual sentiment classification,rnespecially when label data is scarce.