Sara Morrissey

Manual labour: tackling machine translation for sign languages

This article explores the application of data-driven machine translation (MT) to sign languages (SLs). The provision of an SL MT system can facilitate communication between Deaf and hearing people by translating information into the native and preferred language of the individual. In this paper we address data-driven SL MT predominantly for Irish SL (ISL) but also for German SL (DGS/Deutsche Gebärdensprache). We take two different purpose-built corpora to feed our MaTrEx MT system and in a set of experiments translating both to and from the SLs, we investigate the effects of SL data on statistical MT (SMT). Exploiting the bidirectionality of the MaTrEx system, we demonstrate how additional modules, such as recognition and SL animation, can potentially build a full SL MT model for spoken and SL communication in addition to promising evaluation scores. A secondary focus of the article is on the two main issues affecting SL MT, those of transcription and evaluation. We offer a discussion on both these common problems before concluding.

Yorick Wilks, Machine Translation: Its Scope and Limits

Considerable developments, disputes and convergences have befallen machine translation (MT) research in the past 40 years. This book presents the author’s views and the culmination of his research in the field over the course of this period using his own previously published articles as a kind of roadmap through history, with additional afterwords supplementing most chapters bringing previous work into a more modern context. Each chapter is predominantly a snapshot of activities and the author’s views at the time of writing followed by a contemporary commentary. The book is divided into three sections: MT Past, MT Present and MT Future. MT Past (Chapters 2–4) focuses on artificial intelligence (AI) as applied to MT and describes the authors work in the late 60s and 70s including a semantics-based system developed in Stanford in the early 70s. Chapter 4 compares SYSTRAN with IBM approaches and addresses evaluation through the vehicle of his own research as applied to testing SYSTRAN on a new subject area. One might expect MT Present to overview research activities in the recent past, perhaps since the millennium, and major research projects, developments and trends ongoing or emerging. However, Chapters 5–10 are primarily concerned with MT research in the 90s. Take for example the introduction to Chapter 5 where Wilks demonstrates a “revival of interest in the reversibility [of analysis and generation] issue in recent years” by listing papers from 1987 to 1993. This section does present some of the basic paradigms used in MT, for example Chapter 5 discusses analysis and generation as symmetrical opposites and Chapters 6 and 7 deal with interlinguas and statistical approaches, respectively. Chapter 8 addresses the revival of MT in the US in 1990 following US Government interest and funding and appears as a sort of motivational paper, listing strategic and technical reasons for doing MT. Chapters 9 and 10 then