Exploring Distributional Representations and Machine Translation for Aspect-based Cross-lingual Sentiment Classification.

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1613–1623, Osaka, Japan, December 11-17 2016. Exploring Distributional Representations and Machine Translation for Aspect-based Cross-lingual Sentiment Classification Jeremy Barnes Patrik Lambert* Toni Badia Universitat Pompeu Fabra, Barcelona, Spain *Webinterpret, Barcelona, Spain {jeremy.barnes,toni.badia}@upf.edu, patrik.l@webinterpret.com Abstract Cross-lingual sentiment classification (CLSC) seeks to use resources from a source language in order to detect sentiment and classify text in a target language. Almost all research into CLSC has been carried out at sentence and document level, although this level of granularity is often less useful. This paper explores methods for performing aspect-based cross-lingual sentiment classification (aspect-based CLSC) for under-resourced languages. Given the limited nature of parallel data for under-resourced languages, we would like to make the most of this resource for our task. We compare zero-shot learning, bilingual word embeddings, stacked denoising autoen- coder representations and machine translation techniques for aspect-based CLSC. We show that models based on distributed semantics can achieve comparable results to machine translation on aspect-based CLSC. Finally, we give an analysis of the errors found for each method. 1 Introduction Sentiment analysis (SA) seeks to define the underlying sentiment of a text. The best results in SA require the use of a large number of resources; from tokenizers and parsers to large sentiment lexicons or hand- annotated corpora. The creation of these resources requires time, effort and a considerable monetary investment in order to ensure the quality and subsequent usefulness. Therefore, finding a way to perform sentiment analysis for under-resourced languages without having to repeat these efforts is an interesting endeavor. Cross-lingual Sentiment Analysis (CLSA) attempts to find methods to do just this. Most research on CLSA has been at document or sentence level. However, this does not capture the true granularity of opinionated text. Thus, in this paper we focus on CLSA at aspect-level1 Document- and sentence-level CLSA assume that an entire section of text expresses one sentiment towards one entity. This, however, is not always true. Aspect-based CLSA allows for multiple opinions towards multiple entities or aspects. Aspect-based CLSA can be decomposed into three subtasks: en- tity/aspect extraction, opinion holder extraction and sentiment classification. This last task, known as Cross-lingual Sentiment Classification (CLSC), has received little attention at aspect-level. Yet it would greatly benefit companies and government organizations that wish to gather information on the public opinions of their products or policies. In this paper we will only deal with improving or enabling aspect- based CLSC and leave the final goal of creating full aspect-based CLSA systems for under-resourced languages as future work. Most research in CLSC has used Statistical Machine Translation (SMT) as a way of bridging the gap between languages, but there are drawbacks to this. First, an SMT system must be available for the lan- guage combination at hand. This requires a great deal of development and the quality of the sentiment analysis system used afterwards depends heavily on the quality of the SMT system. Secondly, study shows that even high …

Content truncated to 500 words