Cross-lingual sentiment classification: Similarity discovery plus training data adjustment

Abstract

The performance of cross-lingual sentiment classification is sharply limited by the language gap, which means that each language has its own ways to express sentiments. Many methods have been designed to transmit sentiment information across languages by making use of machine translation, parallel corpora, auxiliary unlabeled samples and other resources. In this paper, a new approach is proposed based on the selection of training data, where labeled samples highly similar to the target language are put into the training set. The refined training samples are used to build up an effective cross-lingual sentiment classifier focusing on the target language. The proposed approach contains two major strategies: the aligned-translation topic model and the semi-supervised training data adjustment. The aligned-translation topic model provides a cross-language representation space in which the semi-supervised training data adjustment procedure attempts to select effective training samples to eliminate the negative influence of the semantic distribution differences between the original and target languages. The experiments show that the proposed approach is feasible for cross-language sentiment classification tasks and provides insight into the semantic relationship between two different languages.

Introduction

As Internet access has become globally convenient, our world has experienced the development of a new fashion of social media, in which the volume of user-generated content on the web has massively increased through applications such as Facebook, Twitter, Flickr and LinkedIn, as well as commercial web sites. The value of such unbiased, real-time user-generated content has been shown to be tremendous, with applications in areas such as marketing, decision support systems, politics and public policy support. Due to the enormous amount of user content, it is a very difficult and challenging task to summarize information from online user content. Many natural language processing and information retrieval systems have been designed to automatically treat text and opinion utilizing subjectivity and sentiment analysis [1], [2], [3], [4].

Different methods have been applied in sentiment classification tasks. These methods can be categorized into two main groups: lexicon-based and corpus-based [1], [5], [6]. Both the lexicon-based and corpus-based methods draw sentiment classification information from expert-annotated data sets. However, all of these sentiment classification resources are established in a limited number of languages, which leads to a resource imbalance between different languages. Most sentiment classification resources are written in English [7], [8]. Furthermore, the manual construction of reliable sentiment resources is a very difficult and time-consuming task. Therefore, it would be advantageous to utilize labeled sentiment resources in one language (i.e., English) for sentiment classification in another language. This fantastic idea motivates an interesting research area called cross-lingual sentiment classification (CLSC). The most direct solution to this problem is to use machine translation systems to directly project the information from one language into another [7], [8], [9], [10], [11], [12]. Most existing works in this area have applied machine translation systems to translate labelled training data from the source language into the target language and perform sentiment classification into the target language [13], [14]. Other researchers have employed machine translation in another manner, translating unlabeled test data from the target language into the source language and performing the classification in the source language [7], [11], [15]. A limited number of research works have used both directions of translation to create two different views of the training and test data to compensate for some of the translation limitations [8], [12], [16], [17].

The large gap between different languages occurs naturally. Every language has its unique linguistic terms and writing styles. Even when expressing similar idea, there can be a great disparity in the metaphor and vocabulary in contents of different languages, leading to a much smaller word and phrase intersection between translations and native expressions, as well as different semantic feature distributions between original language and target language contents. As a result, CLSC tasks cannot achieve performance comparable to that obtained for monolingual sentiment classification tasks. To alleviate the problem of this language gap, auxiliary unlabeled corpora or unlabeled parallel corpora are added into the training stage [8], [18] to provide more bilingual word features. This strategy extends the training set and makes origin language and target language closer in the representation space. However, the complementary data brings in useful information as well as noise at the same time because it does not have exactly the same distribution as the training data and the test data. Especially when there is already a distribution disparity between the training set and the test set, noise of the complementary data may cause more adverse impacts than benefits. Additionally, the complementary data itself requires effort to be obtained, which imposes restrictions on the application of these methods.

Based on the above analysis, we try to overcome the difficulty of distribution disparity directly without any auxiliary samples. This paper proposes a novel CLSC strategy of similarity discovery plus training data adjustment (SD-TDA). In the similarity discovery phase, we set up an aligned-translation topic model to generate a bilingual concept representation space where the difference in content between samples from both origin and target languages can be measured through the topic distribution. As the aligned-translation topic model takes in co-occurrence information of terms in one language and between the original and target languages, the relevance of cross-lingual sentiment can be sharply enhanced. Then, in the training data adjustment phase, we set up a semi-supervised process to generate a training set suitable for the target test set. The generated training set is a part of the labeled original language samples that are similar to the reference samples. The reference samples are informative unlabeled target language samples that can be classified by the semi-supervised process with a high degree of confidence. The final classifier will be trained on the generated training set to fulfill the cross-lingual sentiment classification task. When these two steps work together, the sentiment similarity between languages can be maximized, while the distribution gap can be minimized. Generally, our strategy forms a new framework to fit the distribution disparity between the training set and the test set.