The Effect of Comment Errors in Multilingual Electronic Meetings

Abstract

INTRODUCTION The language barrier is a major source of frustration, dissatisfaction, and friction among employees in many multinational organizations (Neal, 1998; Tung & Quaddus, 2002). In an attempt to overcome this communication problem, many multinational companies are increasingly adopting English as a lingua franca (Piekkari, Vaara, Tienari, & Santii, 2005). However, some employees might not be sufficiently skilled in the chosen language to communicate well (Charles & Marschan-Piekkari, 2002; Kim & Bonk, 2002), and those who lack language skills might not be invited to attend key meetings (Louhiala-Salminen, Charles, & Kankaanranta, 2005). Human interpretation can be used as an alternate solution, but it is often expensive or unavailable. In addition, participants in oral meetings must take turns speaking, reducing the amount of individual air time (Karahanna, Ahuja, Srite, & Galvin, 2002). On the other hand, electronic meetings with automatic translation of typed comments can be more efficient and effective (Fugen, Waibel, & Kolss, 2007), and prior studies have shown that comprehension can be relatively good (Aiken, 2008). When poor translation results occurred, these were sometimes attributed to spelling and grammatical errors in the source comments (e.g., Aiken, Martin, Paolillo, & Shirani, 1994), but no study has investigated the relationships among text errors, individual comprehension, ease of use, and system usefulness. From a practical perspective, it is important to know the extent that text errors affect ultimate translation comprehension, and from a theoretical perspective, this research will increase our understanding of the complex relationships among these selected variables. First, we review prior studies of text errors in electronic communication and then describe a recently developed meeting system capable of translating among 58 different languages. Next, we present a model of language intermediation to develop hypotheses relating to textual errors, and an experiment with the multilingual meeting system is performed to test these hypotheses. The paper concludes with experiment limitations and directions for future research. LITERATURE REVIEW Text Errors in Electronic Communication Most people make spelling errors while typing, even when trying to be accurate, and word error rates have varied based upon skill, text difficulty, and time constraints from 2.5% (Mitton, 1996) to 4.75% (Park, 2008). With electronic mail and especially instant messaging, errors can occur much more often as these forms of communication tend to be more informal and hurried, and there might be less concern for accuracy (Bloch, 2002; Jacobs, 2008; Wang & Chang, 1997). For example, one study of email messages showed that narrative sentences contained 2.4% grammatical errors, informative sentences averaged 2.9%, persuasive sentences averaged 3.4%, and expressive sentences averaged 3.5% (Li, 2000). Electronic meetings are also characterized by informal communication, and one study showed about 6% spelling and grammatical errors (Rebman, Aiken, & Cegielski, 2003) while another showed 8.9% (Park, Aiken, Lindblom, & Vanjani, 2010). Misspellings and poor grammar might not cause comprehension to decline much when a single language is used in a meeting, but these errors could compound when translation is involved in a multilingual discussion. An Example Multilingual Electronic Meeting System Multilingual, electronic meeting systems over the past 20 years have supported a few languages (Aiken, 2008), but a recently developed system called Polyglot provides translations among 58 different languages in 3,306 combinations (Aiken & Ghosh, 2009). In addition to the wide variety of languages supported, the accuracy appears to be fairly good overall, but some language combinations are better than others. An automatic evaluation of a subset of the languages used in this meeting system resulted in the BLEU score (0 to 100) averages in Table 1, organized by language group (Aiken & Balan, 2011). …