Learning to Adapt Credible Knowledge in Cross-lingual Sentiment Analysis

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 419–429, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Learning to Adapt Credible Knowledge in Cross-lingual Sentiment Analysis Qiang Chen∗,†, Wenjie Li†,⋆, Yu Lei†, Xule Liu∗, Yanxiang He∗,‡ ∗School of Computer Science, Wuhan University, China †Department of Computing, The Hong Kong Polytechnic University, Hong Kong ⋆Hong Kong Polytechnic University Shenzhen Research Institute, China ‡The State Key Lab of Software Engineering, Wuhan University, China ∗{qchen, xuleliu, yxhe}@whu.edu.cn †{csqchen, cswjli, csylei}@comp.polyu.edu.hk Abstract Cross-lingual sentiment analysis is a task of identifying sentiment polarities of texts in a low-resource language by using sen- timent knowledge in a resource-abundant language. While most existing approaches are driven by transfer learning, their performance does not reach to a promising level due to the transferred errors. In this paper, we propose to integrate into knowl- edge transfer a knowledge validation mod- el, which aims to prevent the negative influence from the wrong knowledge by distinguishing highly credible knowledge. Experiment results demonstrate the neces- sity and effectiveness of the model. 1 Introduction With the wide range of business value, sentiment analysis has drawn increasing attention in the past years. The extensive research and development efforts produce a variety of reliable sentiment resources for English, one of the most popular language in the world. These available rich resources become the treasure of knowledge to help conduct or enhance sentiment analysis in the other languages, which is a task known as cross-lingual sentiment analysis (CLSA). In the literature of CSLA, the language with abundant reliable resources is called the source language (e.g., English), while the low-resource language is referred to as the target language (e.g., Chinese). However, in this paper, the situation is a low resource language scenario, where the source language is English, and the target language is Chinese. The main idea of existing CLSA researches is to first build up the connection between the source and target languages to overcome the language barrier, and then develop an appropriate knowl- edge transfer approach to leverage the annotated data from the source language to train a sentiment classification model in the target language, either supervised or semi-supervised. In particular, these approaches exploit and convert the knowledge learned from the source language to automatically generate and expand the pseudo-training data for the target language. The machine translation (MT) service is one of the most common ways used to build the language connection (Wan, 2008; Banea et al., 2008; Wan, 2009; Wei and Pal, 2010; Gui et al., 2014). Although it is claimed in Duh et al. (2011) that the MT service is ripe for CLSA, the imperfect MT quality hinders existing MT- based CLSA approaches from the further advance. In our preliminary study, we find that even the Google translator1 (i.e., one of the most widely used online MT service (Shankland 2013)) may unavoidably changes the sentiment polarity of the translated text, as illustrated below, with a …

Content truncated to 500 words