The Impact of Indirect Machine Translation on Sentiment Classification

The Impact of Indirect Machine Translation on Sentiment Classification Alberto Poncelas alberto.poncelas@adaptcentre.ie Pintu Lohar pintu.lohar@adaptcentre.ie Andy Way andy.way@adaptcentre.ie ADAPT Centre, Dublin City University, Glasnevin, Dublin 9, Ireland James Hadley hadleyj@tcd.ie Trinity Centre for Literary and Cultural Translation and ADAPT Centre, Trinity College Dublin, Ireland Abstract Sentiment classification has been crucial for many natural language processing (NLP) applica- tions, such as the analysis of movie reviews, tweets, or customer feedback. A sufficiently large amount of data is required to build a robust sentiment classification system. However, such resources are not always available for all domains or for all languages. In this work, we propose employing a machine translation (MT) system to translate customer feedback into another language to investigate in which cases translated sentences can have a positive or negative impact on an automatic sentiment classifier. Furthermore, as performing a direct translation is not always possible, we explore the performance of automatic classifiers on sentences that have been translated using a pivot MT system. We conduct several experiments using the above approaches to analyse the performance of our proposed sentiment classification system and discuss the advantages and drawbacks of classifying translated sentences. 1 Introduction The key factor for building a sentiment classifier is training the model using a dataset of at least two different sentiment classes. This process requires a huge amount of data to build a workable sentiment classification system. However, sometimes it is difficult to find the required resources for the pertinent domain in large enough quantities and in all relevant languages. A typical and efficient approach to solve this problem consists in building the sentiment classification system with a high-resource language such as English and translating the sentences to be classified into this high-resource language. An example of this is customer feedback classification. Many companies face difficul- ties analyzing what their clients say about their products and/or services because of the vast amount of feedback. This problem is aggravated for those businesses that have clients from several countries and therefore receive feedback in multiple languages. Although the feed- back in multiple languages can be translated into English, the translated sentences may contain errors. Moreover, as the feedback consists of user-generated texts, it tends to be less grammati- cally strict than texts from literature, which implies that finding an accurate translation becomes more difficult (Lohar et al., 2019). Proceedings of the 14th Conference of the Association for Machine Translation in the Americas October 6 - 9, 2020, Volume 1: MT Research Track Page 78 MT models are built to generate translations that carry the same meaning as the original sentence and are also fluent in the target language. However, in some scenarios, metrics for measuring fluency are less relevant. This is the case in the sentiment analysis of translated sentences, where maintaining the sentiment is the priority, even if the translation is not accurate in terms of adequacy and fluency (Tebbifakhr et al., 2019). Despite the MT system generating understandable translations, it may not manifest the same sentiment as …

Content truncated to 500 words