Sentiment Preservation in Review Translation using Curriculum-based Re-inforcement Framework

Sentiment Preservation in Review Translation using Curriculum-based Re-inforcement Framework Divya Kumari∗, Soumya Chennabasavraj∗∗, Nikesh Garera∗∗ and Asif Ekbal∗ ∗Department of Computer Science and Engineering, Indian Institute of Technology Patna, Patna, India ∗∗Flipkart, India ∗divya_1921cs10,asif@iitp.ac.in ∗∗soumya.cb,nikesh.garera@flipkart.com Abstract Machine Translation (MT) is a common approach to feed humans or machines in a cross-lingual context. However, there are some expected drawbacks. Studies suggest that in the cross- lingual context, MT system often fails to preserve different stylistic and pragmatic properties of the source text (e.g. sentiment, emotion, gender traits, etc.) to the target translation. These disadvantages can degrade the performance of any downstream Natural Language Processing (NLP) applications, such as sentiment analyser, that heavily relies on the output of the MT systems (especially in a low-resource setting). The susceptibility to sentiment polarity loss becomes even more severe when an MT system is employed for translating a source content that lacks a legitimate language structure (e.g. review text). Therefore, while improving the general quality of the Neural Machine Translation (NMT) output (e.g. adequacy), we must also find ways to minimize the sentiment loss in translation. In our current work, we present a deep re-inforcement learning (RL) framework in conjunction with the curriculum learning to fine-tune the parameters of a full-fledged NMT system so that the generated translation successfully encodes the underlying sentiment of the source without compromising the adequacy, unlike the previous method. We evaluate our proposed method on the English–Hindi (product domain) and French–English (restaurant domain) review datasets, and found that our method (further) brings a significant improvement over a full-fledged supervised baseline for the machine translation and sentiment classification tasks. 1 Introduction Product and/or service reviews available in the e-commerce portals are predominantly in the English language, and hence a large number of population can not understand these. Machine Translation (MT) system can play a crucial role in bridging this gap by translating the user- generated contents, and directly displaying them, or making these available for the downstream Natural Language Processing (NLP) tasks e.g. sentiment classification1 (Araújo et al., 2020; Barnes et al., 2016; Mohammad et al., 2016; Kanayama et al., 2004). However, numerous studies (Poncelas et al., 2020a; Afli et al., 2017; Mohammad et al., 2016; Sennrich et al., 2016a) have found a significant loss of sentiment during the automatic translation of the source text. The susceptibility to sentiment loss aggravates when the MT system is translating a noisy review that lacks a legitimate language structure at the origin. For example, a noisy review con- tains several peculiarities and informal structures, such as shortening of words (e.g. “awesome” as “awsm”), acronyms (e.g. “Oh My God” as “OMG”), phonetic substitution of numbers (e.g. “before” as “b4”), emphasis on characters to define extremity of the emotion (e.g. “good” as “gooooooood”), spelling mistakes, etc. Unfortunately, even a pervasively used commercial neural 1Please note that our current work is limited to cross-lingual sentiment analysis [CLSA] via MT based approach. Proceedings of the 18th Biennial Machine Translation Summit Virtual USA, August 16 - 20, 2021, Volume …

Content truncated to 500 words