Towards a Linguistically Motivated Segmentation for a Simultaneous Interpretation System

Towards a Linguistically Motivated Segmentation for a Simultaneous Interpretation System Youngeun Koo1, Jiyoun Kim1, Jungpyo Hong1, Munpyo Hong1* and Sung-Kwon Choi2 1Dept. of German Linguistics & Literature, Sungkyunkwan University, 25-2, Sungkyunkwan-ro, Jongno-gu, Seoul, Korea 2Language Intelligence Research Section, Electronics and Telecommunications Research Institute(ETRI), 218 Gajeong-ro, Yuseong-gu, Daejeon, Korea {sarah8835,kite92,jphong2800,skkhmp}@skku.edu;choisk@etri.re.kr Abstract For simultaneous interpretation, it is very important to identify appropriate segmentation boundaries so that the source text can be translated accurately and promptly. This paper proposes four different segmentation methods for a simultaneous interpretation system. These methods are designed considering the balance between translation accuracy and translation latency. They employ various linguistic features such as prosodic, part-of-speech (POS), dependency, discourse, and cognitive features. This paper conducts experiments on segmentation in English to Korean and Korean to English simultaneous interpretation. Our finding shows that different segmentation method should be applied depending on the source language. 1 Introduction Simultaneous interpretation aims to accurately translate what is being said in a source language into a target language quickly. To this aim, a strategy that segments the source text at appropriate points is often used by both human interpreters and simultaneous interpretation systems. Ideally, simultaneous interpretation systems should provide interpretation results as soon as possible while minimizing translation latency. However, there is a trade-off between translation accuracy and latency. The longer the segmentation unit is, the higher the translation accuracy will be, while the latency gets worse. In contrast, if the segmented unit is short, the  Corresponding author latency will be better; however, the accuracy tends to be worsened. In this paper, we investigate various segmentation features to determine the optimal segmentation points. The features were designed through a linguistic investigation into the prosodic, POS, dependency, and discourse-level characteristics in simultaneous interpretation. Also, we tried to find out what the appropriate segmentation length should be. In this paper, we propose four methods to derive optimal segmentation points employing these features. The segmentation features and methods considering both translation accuracy and latency may help to improve the performance of a simultaneous interpretation system. In section 2, related works are introduced. In section 3 we suggest linguistic features for the segmentation. Section 4 shows our experimental setup with proposed features and methods and analyses the results. Finally, Section 5 concludes this paper and discusses future researches. 2 Related Works Previous researches showed various approaches for investigating segmentation boundaries. The simplest way is to find a possible sentence unit (Cettolo and Federico, 2006; Sridhar et al., 2013b). Sridhar et al. (2013b) found segmentation boundaries based on predicting possible sentence end. Also, Sridhar et al. (2013b) utilized commas in sentences for segmentation. Another approach for segmentation boundary detection is to use POS of the source text (Stolcke and Shriberg, 1996; Sridhar et al., 2013b; Nakabayashi et al., 2019). Stolcke and Shriberg (1996) tested two models for segmentation based on the POS of an input. The first model used POS tags labeled on every token and the second model used both POS and ‘segmentation related’ information, such as …

Content truncated to 500 words