Research On Cross-Lingual Aspect-Level Sentiment Classification

The aspect-level sentiment classification task aims to judge the sentiment tendency corresponding to a certain aspect in a sentence.In recent years,it has attracted extensive attention of scholars in the field of natural language processing.Since most of the existing related research is based on English corpus,the distribution of sentiment resources in different languages is not balanced,and cross-lingual methods emerge as the times require.Cross-lingual aspect-level sentiment classification aims to use the resources in the source language to help the target language to perform aspect-level sentiment detection and classification,and its core problem is how to realize the effective sharing of cross-lingual knowledge.This paper conducts research on cross-lingual aspect-level sentiment classification to solve this problem.The specific research contents are as follows:(1)To solve the problem of semantic association between texts in different languages,this paper proposes a cross-lingual aspect-level sentiment classification method based on attention mechanism.First,the self-attention mechanism is used to learn the own features of the text;then,a cross-lingual attention mechanism is introduced to mine the association features of bilingual text pairs;finally,an aspect-specific attention mechanism is used to capture the contextual features about the aspects,obtain the sentence representations that incorporate multi-sentiment,cross-lingual information.Experiments show that the attention mechanism-based method effectively constructs the relationship between cross-lingual texts and significantly improves the accuracy of cross-lingual aspect-level sentiment classification.(2)To solve the problem of the global dependency structure of cross-lingual aspects and sentences,this paper proposes a cross-lingual aspect-level sentiment classification method based on graph convolutional networks.First,a cross-lingual aspect-sentence heterogeneous graph is constructed using the entire corpus,and sentences and aspects are regarded as graph nodes;then,related nodes are connected through sentence translation relationship,aspect translation relationship and sentence-aspect co-occurrence relationship to capture global dependency structure information of the cross-lingual aspects and sentences;finally,the graph structure and node encoding are fed into a graph convolutional network.Experiments show that the method based on graph convolutional network can effectively extract and integrate heterogeneous information in bilingual corpora,and help improve the performance of cross-lingual aspect-level sentiment classification.(3)To solve the problem of limited annotation samples for aspect-level sentiment classification,this paper proposes a cross-lingual aspect-level sentiment classification method based on multi-task learning.First,extract large-scale document-level labeled data as an auxiliary training set;then,construct two cross-lingual auxiliary tasks on this data set:cross-lingual document-level sentiment classification task and bilingual synonym judgment task;finally,the main and auxiliary tasks are jointly learned to fine-tune the M-BERT model.Experiments show that the method based on multi-task learning can effectively transfer the semantic-grammatical knowledge shared by document-level texts and aspect-level texts and capture cross-lingual correlation information,thereby improving the performance of crosslingual aspect-level sentiment classification models.