Statistical Analysis of Missing Translation in Simultaneous Interpretation Using A Large-scale Bilingual Speech Corpus

Statistical Analyses of Missing Translations in Simultaneous Interpretation Using a Large-scale Bilingual Speech Corpus Zhongxi Cai1, Koichiro Ryu2, Shigeki Matsubara3 1Graduate School of Informatics, Nagoya University, 2Tsu City College 3Information and Communications, Nagoya University 1,3Furo-cho, Chikusa-ku, Nagoya, 464-8601, Japan 2Isshindennakano 157, Tsu, Mie, 514-0112, Japan cai.zhongxi@f.mbox.nagoya-u.ac.jp, ryu@tsu-cc.ac.jp, matubara@nagoya-u.jp Abstract This paper describes statistical analyses of missing translations in simultaneous interpretations. Eighty-eight lectures from English-to- Japanese interpretation data from a large-scale bilingual speech corpus were used for the analyses. Word-level alignment was provided manually, and English words without corresponding Japanese words were considered missing translations. The English lectures con- tained 46,568 content words, 33.1% of which were missing in the translation. We analyzed the relationship between missing translations and various factors, including the speech rate of the source language, delay of interpretation, part-of-speech, and depth in the syntactic structure of the source language. The analyses revealed that the proportion of missing translations is high when the speech rate is high and delay is large. We also found that a high proportion of adverbs were missed in the translations, and that words at deeper positions in the syntactic structure were more likely to be missed. Keywords: Simultaneous interpretation, Speech corpus, Bilingual corpus, Speech translation 1. Introduction In simultaneous interpretation, an interpreter must convey their translation to the target language while simultaneously listening to, comprehending, and memorizing the content of the source language speech. This is challenging and in- volves a number of difficulties that can reduce the quality of the interpretation. If such difficulties can be detected auto- matically, a support environment for simultaneous interpre- tation that presents the speech content or provides candidate translations could be achieved. Gile (1995) referred to interpretation difficulties as “prob- lem triggers”. Such triggers include the density of the source speech (e.g., the delivery rate and the density of in- formation in the content). In addition, unfamiliar names, numbers, and complex syntactic structures, etc. are chal- lenging for interpreters. Since most of the time interpreters work at near capacity, such additional challenges can lead to failure (Gile, 1999). Interpretation failures include er- rors, omissions, and infelicities (Gile, 2009). In this paper, omissions are used to identify when an interpreter has en- countered a difficulty. Time constraints are severe in simultaneous interpretation such that it is impossible to translate all speech content; thus, omissions are inevitable (Dillinger, 1994). To im- prove interpretation quality or develop a method to train interpretation skills, several studies have investigated omis- sions in simultaneous interpretation. Various types of omis- sions have been described and several factors related to omissions have been studied (Barik, 1994). However, these analyses, which were based on observation, did not clarify the correlation between the identified factors and the occur- rence of omissions. In this paper, to detect occasions where interpreters would encounter difficulties, we statistically analyzed the correla- tion between source speech features and interpretation con- ditions and the occurrence of missing translations. In the analyses, we used 88 lectures of English-to-Japanese (E- J) interpretation data from …

Content truncated to 500 words