Ha, Le An
AffiliationFaculty of Science and Engineering
MetadataShow full item record
AbstractEmotions are biological states of feeling that humans may verbally express to communicate their negative or positive mood, influence others, or even afflict harm. Although emotions such as anger, happiness, affection, or fear are supposedly universal experiences, the lingual realisation of the emotional experience may vary in subtle ways across different languages. For this reason, preserving the original sentiment of the source text has always been a challenging task that draws in a translator's competence and fi nesse. In the professional translation industry, an incorrect translation of the sentiment-carrying lexicon is considered a critical error as it can be either misleading or in some cases harmful since it misses the fundamental aspect of the source text, i.e. the author's sentiment. Since the advent of Neural Machine Translation (NMT), there has been a tremendous improvement in the quality of automatic translation. This has lead to an extensive use of NMT online tools to translate User-Generated Text (UGT) such as reviews, tweets, and social media posts, where the main message is often the author's positive or negative attitude towards an entity. In such scenarios, the process of translating the user's sentiment is entirely automatic with no human intervention, neither for post-editing nor for accuracy checking. However, NMT output still lacks accuracy in some low-resource languages and sometimes makes critical translation errors that may not only distort the sentiment but at times flips the polarity of the source text to its exact opposite. In this thesis, we tackle the translation of sentiment in UGT by NMT systems from two perspectives: analytical and experimental. First, the analytical approach introduces a list of linguistic features that can lead to a mistranslation of ne-grained emotions between different language pairs in the UGT domain. It also presents an error-typology specifi c to Arabic UGT illustrating the main linguistic phenomena that can cause mistranslation of sentiment polarity when translating Arabic UGT into English by NMT systems. Second, the experimental approach attempts to improve the translation of sentiment by addressing some of the linguistic challenges identifi ed in the analysis as causing mistranslation of sentiment both on the word-level and on the sentence-level. On the word-level, we propose a Transformer NMT model trained on a sentiment-oriented vector space model (VSM) of UGT data that is capable of translating the correct sentiment polarity of challenging contronyms. On the sentence-level, we propose a semi-supervised approach to overcome the problem of translating sentiment expressed by dialectical language in UGT data. We take the translation of dialectical Arabic UGT into English as a case study. Our semi-supervised AR-EN NMT model shows improved performance over the online MT Twitter tool in translating dialectical Arabic UGT not only in terms of translation quality but also in the preservation of the sentiment polarity of the source text. The experimental section also presents an empirical method to quantify the notion of sentiment transfer by an MT system and, more concretely, to modify automatic metrics such that its MT ranking comes closer to a human judgement of a poor or good translation of sentiment.
CitationSaadany, H. (2022) A study of the translation of sentiment in user-generated text. Wolverhampton: University of Wolverhampton. http://hdl.handle.net/2436/625110
PublisherUniversity of Wolverhampton
TypeThesis or dissertation
DescriptionA thesis submitted in partial ful filment of the requirements of the University of Wolverhampton for the degree of Doctor of Philosophy.
The following licence applies to the copyright and re-use of this item:
- Creative Commons
Except where otherwise noted, this item's license is described as Attribution-NonCommercial-NoDerivatives 4.0 International