Deep Learning Paradigm with Transformed Monolingual Word Embeddings for Multilingual Sentiment Analysis

Deep Learning Paradigm with Transformed Monolingual Word Embeddings for Multilingual Sentiment Analysis Yujie Lu ZhongAn Technology Data Science Lab luyujie@zhongan.io Boyi Ni ZhongAn Technology Data Science Lab niboyi@zhongan.io Qijin Ji ZhongAn Technology Data Science Lab jiqijin@zhongan.io Kotaro Sakamoto Yokohama National University sakamoto@forest.eis.ynu.ac.jp Hideyuki Shibuki Yokohama National University shib@forest.eis.ynu.ac.jp Tatsunori Mori Yokohama National University mori@forest.eis.ynu.ac.jp Abstract The surge of social media use has triggered huge demands of multilingual sentiment anal- ysis (MSA) for various purposes, such as unveiling cultural difference. So far, tradi- tional methods resorted to machine transla- tion (MT)—translating other languages to En- glish, then adopted the existing methods of English. However, this paradigm is highly conditioned by the quality of MT. In this pa- per, we propose a new deep learning paradigm for MSA that assimilates the differences be- tween languages. First, separately pre-trained monolingual word embeddings in different spaces are mapped into a shared embedding space; then, a parameter-sharing deep neu- ral network using those mapped word em- beddings for MSA is built. The experimen- tal results justify the effectiveness of the pro- posed paradigm. Especially, our convolutional neural network (CNN) model with orthogo- nally mapped word embeddings outperforms a state-of-the-art baseline by 3.4% in terms of classification accuracy. 1 Introduction The prevalence of social media has allowed the col- lection of abundant subjective multilingual texts. Twitter is such a particularly significant multilingual data source providing researchers/companies with sufficient opinion pieces on various topics/products from all over the world. By analyzing these mul- tilingual opinion texts, there can be many useful applications, such as revealing the cultural varia- tions and conducting customer surveys in different areas. Therefore, it’s necessary to develop an effec- tive MSA model that can process multilingual texts simultaneously. The research on MSA has progressed slowly com- pared with monolingual sentiment analysis, mainly due to the lack of a benchmark dataset that can assess the cross-language adaptability of methods. As many previous studies have highlighted, open- source sentiment datasets are usually imbalanced between different languages (Mihalcea et al., 2007; Denecke, 2008; Wan, 2009; Steinberger et al., 2011). There are many freely available annotated sentiment corpora for English, but not for many other languages. As a compromise, many of the pre- vious multilingual corpora have been built using hu- man/machine translation, which are not authentic. In this study, we used the MDSU corpus as our training/test dataset (Lu et al., 2017)1. The MDSU corpus with 5,422 tweets in total contains three languages, i.e., English, Japanese, and Chinese, and involves four identical international topics, i.e., iPhone 6, Windows 8, Vladimir Putin, and Scottish Independence. The multilinguality and topic distri- bution of the corpus makes it an ideal MSA dataset. Moreover, monolingual sentiment analysis meth- ods are usually not portable between languages, 1The corpus can be downloaded from https://github.com/lyjlyj517/The-MDSU-Corpus. PACLIC 32 419 32nd Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018 Copyright 2018 by the authors since they depend on language-specific polarity lex- icons, POS taggers and parsers, etc. This prevents the …

Content truncated to 500 words