Real-Time Spoken Language Translation Using Associative Processors

Real-Time Spoken Language Translation Using Associative Processors Kozo Oi, Eiichiro Sumita, Osamu Furuse, Hitoshi Iida and Tetsuya Higuchi~ ATR Interpreting Telecommunications Research Laboratories 2-2 Hikaridai, Seika, Souraku, Kyoto 619-02, JAPAN { o i, sumit a, furus e, i ida} @it i. atr. co. j p ~Electrotechnical Laboratory 1-1-4 Umezono, Tsukuba, Ibaraki 305, Japan higuchi©et i. go. jp Abstract This paper proposes a model using associative processors (APs) for real-time spoken language translation. Spoken language translation re- quires (1) an accurate translation and (2) a real- time response. We have already proposed a model, TDMT (Transfer-Driven Machine Trans- lation), that translates a sentence utilizing ex- amples effectively and performs accurate struc- tural disambiguation and target word selection. This paper will concentrate on the second re- quirement. In TDMT, example-retrieval (ER), i.e., retrieving examples most similar to an in- put expression, is the most dominant part of the total processing time. Our study has concluded that we only need to implement the ER for ex- pressions including a frequent word on APs. Ex- perimental results show that the ER can be dras- tically speeded up. Moreover, a study on com- munications between APs demonstrates the scal- ability against vocabulary size by extrapolation. Thus, our model, TDMT on APs, meets the vital requirements of spoken language translation. 1 Introduction Research on speech translation that began in the mid-1980s has been challenging. Such research has resulted in several prototype systems (Morimoto et al., 1993; Kitano, 1991; Waibel et al., 1991). Speech translation consists of a sequence of processes, i.e., speech recognition, spoken language translation and speech synthesis. Each process must be accelerated in order to achieve real-time response. This pa- per focuses on the second process, spoken language translation, which requires (1) an accurate trans- lation and (2) a real-time response. We have al- ready proposed a model that utilizes examples and translates a sentence by combining pieces of trans- fer knowledge, i.e., target language expressions that correspond to source language expressions that cover the sentence jointly. The model is called Transfer- Driven Machine Translation (TDMT) (Furuse and Iida, 1992; Furuse et al., 1994) (see subsection 2.1 for details). A prototype system of TDMT which trans- lates a Japanese spoken sentence into English, has performed accurate structural disambiguation and target word selection 1. This paper will focus on the second requirement. First, we will outline TDMT and analyze its com- putational cost. Second, we will describe the con- figuration, experimental results and scalability of TDMT on associative processors (APs). Finally, we will touch on related works and conclude. 2 TDMT and its Cost Analysis 2.1 Outline of TDMT In TDMT, transfer knowledge is the primary knowl- edge, which is described by an example-based frame- work (Nagao, 1984). A piece of transfer knowledge describes the correspondence between source lan- guage expressions (SEs) and target language expres- sions (TEs) as follows, to preserve the translational equivalence: SE => TEl (Era, E12,...), : TE, (E,1, E,2,...) Eij indicates the j-th example of TEi. For exam- ple, the …

Content truncated to 500 words