Comparing MT Approaches for Text Normalization

Proceedings of Recent Advances in Natural Language Processing, pages 740–749, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_086 740 Comparing MT Approaches for Text Normalization Claudia Matos Veliz, Orph´ee De Clercq and V´eronique Hoste LT 3, Language and Translation Technology Team - Ghent Univeristy Groot-Brittanni¨elaan 45, 9000, Ghent, Belgium Firstname.Lastname@UGent.be Abstract One of the main characteristics of social media data is the use of non-standard lan- guage. Since NLP tools have been trained on traditional text material, their perfor- mance drops when applied to social me- dia data. One way to overcome this is to first perform text normalization. In this work, we apply text normalization to noisy English and Dutch text coming from different genres: text messages, mes- sage board posts and tweets. We con- sider the normalization task as a Machine Translation problem and test the two lead- ing paradigms: statistical and neural ma- chine translation. For SMT we explore the added value of varying background cor- pora for training the language model. For NMT we have a look at data augmentation since the parallel datasets we are work- ing with are limited in size. Our results reveal that when relying on SMT to per- form the normalization, it is beneficial to use a background corpus that is close to the genre to be normalized. Regarding NMT, we find that the translations - or normalizations - coming out of this model are far from perfect and that for a low- resource language like Dutch adding ad- ditional training data works better than ar- tificially augmenting the data. 1 Introduction Probably one of the most persistent characteris- tics of social media texts is that they are full of non-standard words (Eisenstein, 2013). Several sources of noise can influence the way people write. For example, the different kind of social media platforms available nowadays provide a di- verse range of ways to communicate and particular forms of language variations (Ke et al., 2008). So- cial variables such as gender, age and race can also influence communication style (Schwartz et al., 2013; Blodgett et al., 2016). Location is also an important variable, since it can lead to the use of dialect and non-standard words. (Vandekerckhove and Nobels, 2010; Blodgett et al., 2016). Very typical for this User-Generated Con- tent (UGC) is the expression of emotions by the use of symbols or lexical variations of the words (Van Hee et al., 2017). This can be done in the form of flooding or the repetition of characters, capitalization, and the productive use of emoticons. In addition, the use of ho- mophonous graphemic variants of a word, abbre- viations, spelling mistakes or letter transpositions are very typical, since people tend to write as they speak and/or write as fast as possible. One can imagine that all those characteristics contribute to an increased difficulty of automati- cally processing and analyzing UGC. Since Nat- ural Language Processing (NLP) tools have orig- inally been developed for and trained on standard language, these non-standard forms adversely af- fect language analysis …

Content truncated to 500 words