Sentiment Analysis Across Languages: Evaluation Before and After Machine Translation to English

TL;DR

The paper evaluates sentiment analysis across multiple languages by comparing native-language transformer models (BERT variants) and multilingual models (XLM-RoBERTa) on original text and on English translations produced with OPUS-MT, using the Multilingual Amazon Reviews Corpus. It employs a robust pipeline and two tasks (sentiment classification and star rating) with F1 as the evaluation metric, under constrained compute by using language-specific subsets and a translated English-augmented dataset. The results show that translation preserves sentiment better for European languages than for Asian languages, with XLM-RoBERTa generally outperforming BERT and translation-induced degradations highlighting the need for improved translation and cross-lingual adaptation. The work suggests future directions including sentiment-focused MT fine-tuning, multimodal sentiment cues, and transfer learning to bolster performance for low-resource languages, and provides code for reproducibility.

Abstract

People communicate in more than 7,000 languages around the world, with around 780 languages spoken in India alone. Despite this linguistic diversity, research on Sentiment Analysis has predominantly focused on English text data, resulting in a disproportionate availability of sentiment resources for English. This paper examines the performance of transformer models in Sentiment Analysis tasks across multilingual datasets and text that has undergone machine translation. By comparing the effectiveness of these models in different linguistic contexts, we gain insights into their performance variations and potential implications for sentiment analysis across diverse languages. We also discuss the shortcomings and potential for future work towards the end.

Figures (2)

• Figure 1: Training Pipeline
• Figure 2: Distribution of Labels