Statistical Machine Translation of Texts with Misspelled Words

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 412–419, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Statistical Machine Translation of Texts with Misspelled Words Nicola Bertoldi Mauro Cettolo Marcello Federico FBK - Fondazione Bruno Kessler via Sommarive 18 - 38123 Povo, Trento, Italy {bertoldi,cettolo,federico}@fbk.eu Abstract This paper investigates the impact of mis- spelled words in statistical machine transla- tion and proposes an extension of the transla- tion engine for handling misspellings. The en- hanced system decodes a word-based confu- sion network representing spelling variations of the input text. We present extensive experimental results on two translation tasks of increasing complex- ity which show how misspellings of different types do affect performance of a statistical ma- chine translation decoder and to what extent our enhanced system is able to recover from such errors. 1 Introduction With the widespread adoption of the Internet, of modern communication, multimedia and mobile de- vice technologies, the amount of multilingual in- formation distributed and available to anyone, any- where, has exploded. So called social media have rapidly reshaped information exchange among Inter- net users, providing new means of communication (blogs, tweets, etc.), collaboration (e.g. wikis), and sharing of multimedia content, and entertainment. In particular, social media have today become also an important market for advertisement as well as a global forum for consumer opinions (Kushal et al., 2003). The growing spread of user-generated content is scaling-up the potential demand for on-line machine translation (MT) but also setting new challenges to the field of natural language processing (NLP) in general. The language written and spoken in the social media presents an impressive variety of con- tent and styles (Schler et al., 2006), and writing con- ventions that rapidly evolve over time. Moreover, much of the content is expressed in informal style, that more or less violates the standard grammar, con- tains many abbreviations and acronyms, and finally many misspelled words. From the point of view of MT, language of social media is hence very different from the one represented in the text corpora nowa- days available to train statistical MT systems. Facing all these challenges, we pragmatically scaled down our ambition and decided to investigate a basic, somehow preliminary, well defined prob- lem: the impact of misspelled words in statistical MT. Unintentional typing errors are indeed remark- ably frequent in online chats, blogs, wikis, reviews, and hence constitute a major source of noise (Subra- maniam et al., 2009). In this paper we aim at studying performance degradation of statistical MT under different levels and kinds of noise, and at analyzing to what extent statistical MT is able to recover from errors by en- riching its input with spelling variations. After a brief overview of NLP literature related to noisy texts, in Section 3 we consider different types of misspellings and derive simple but realistic models that are able to reproduce them. Such mod- els are then used to generate errors in texts passed to a phrase-based …

Content truncated to 500 words