A Language-Agnostic Framework with Bidirectional Syntactic Graph Convolutional Networks for Cross-Lingual Aspect Term Extraction

Abstract

Aspect term extraction is a vital sub-task of sentiment analysis, which aims to extract explicit product attributes in customer reviews. Unfortunately, many languages lack sufficient labeled data, so researchers focus on Cross-lingual Aspect Term Extraction (XATE) to fully use sufficient data in other languages. Most recent cross-lingual methods focus on semantic alignment and data augmentation, but lack research on language structure, including syntax and lexicality. To this end, we propose a Language-Agnostic framework with Bidirectional Syntactic Graph Convolutional Networks (LA-BSGCN) for XATE. It is based on the idea that the topological structures of syntactic dependencies and the lexical tags across different languages are similar. We design a multi-layer bidirectional GCN, which can encode the syntactic tree more accurately. Furthermore, to reduce the lexicality semantic gap between different languages, we encode named entity recognition (NER) and part of speech (POS) information into our model. We conduct six pairs of cross-lingual experiments on SemEval2016 Task5 datasets. The results show that our LA-BSGCN significantly reduces the semantic gap and outperforms the state-of-the-art methods. For reproducibility, our code for this paper is available at github.