A New Two-Layer Approach for Spoken Language Translation

A New Two-Layer Approach for Spoken Language Translation Jhing-Fa Wang, Shun-Chieh Lin, and Hsueh-Wei Yang Department of Electrical Engineering, National Cheng Kung University wangjf@csie.ncku.edu.tw Abstract. This study proposes a new two-layer approach for spoken language translation. First, we develop translated examples and transform them into speech signals. Second, to properly retrieve a translated example by analyzing speech signals, we expand the translated example into two layers: an intention layer and an object layer. The intention layer is used to examine intention similarity between the speech input and the translated example. The object layer is used to identify the objective components of the examined intention. Experiments were conducted with the languages of Chinese and English. The results revealed that our proposed approach achieves about 86% and 76% understandable translation rate for the Chinese-to-English and the English-to-Chinese translations, respectively. 1 Introduction With the growing of globalization, people now often meet and do business with those who speak different languages, on-demand spoken language translation (SLT) has become increasingly important (See JANUS III [6], Verbmobil [9], EUTRANS [3] and ATR-MATRIX [1]). Currently, there are two main architectures of SLT: conventional sequential architecture and fully integrated architecture [1]. For the sequential architecture, a spoken language translation is composed by a speech recognition system followed by a linguistic (or non-linguistic) text-to-text translation system. In the integrated architecture, acoustic-phonetic models are integrated into translation models in the similar way as for speech recognition. Recently, an integrated architecture based on stochastic finite-state transducer (SFST) has been presented in [3,4]. The SFST approach integrated three models in a single network where the search process takes place. The three models are Hidden Markov Models for the acoustic part, language models for the source language and finite state transducers for the transfer between the source and target language. The output of this search process is the target word sequence associated to the optimal path. Fig. 1 shows an example of the SFST approach. λ denotes the empty string. The source sentence “una habitación doble” can be translated to either “a double room” or “a room with two beds”. The most probable translation is the first one with probability of 0.09. Fig. 1. Examples of the stochastic finite-state transducer However, when the training data of SFST is insufficient, the results obtained by the sequential architecture are better than the results obtained by the integrated architecture [4]. In addition, word reordering is still a thorny problem in SFST which is based on statistical-based translation methods [5]. Therefore, we propose adopting example-based approaches for better integration. Such the adopted approach does not require the database to be as large as in SFST and can utilize word mappings between source-target language of a chosen translated example for word reordering [2,8]. In this paper, we further propose a new two-layer approach for the example-based spoken language translation. First, we develop translated examples and transform them into speech signals. Second, to properly retrieve a translated example by analyzing speech signals, we expand the translated example into two …

Content truncated to 500 words