Text and speech translation by means of subsequential transducers

Abstract

The full paper explores the possibility of using Subsequential Transducers (SST), a finite state model, in limited domain translation tasks, both for text and speech input. A distinctive advantage of SSTs is that they can be efficiently learned from sets of input-output examples by means of OSTIA, the Onward Subsequential Transducer Inference Algorithm (Oncina et al. 1993). In this work a technique is proposed to increase the performance of OSTIA by reducing the asynchrony between the input and output sentences, the use of error correcting parsing to increase the robustness of the models is explored, and an integrated architecture for speech input translation by means of SSTs is described.

Learning finite-state models for machine translation

“…It allowed, for instance, to understand the need for source and/or target language models in order to cope with insufficient training data and/or imperfect input text. All in all, very good results (TWER below 1%) were achieved for the MLT corpus using OSTIA-DR along error-correcting smoothing (Vilar, Vidal, & Amengual, 1996;Vilar et al, 1996a). The application of OSTIA and OSTIA-DR to the more realistic EUTRANS task was first described in Vidal (1997).…”

“…For the experiments reviewed in the next section, 8K SpanishEnglish sentence pairs (about 120K running words per language) were used for training. Testing was performed on another 10K sentences different from those used in training (Vilar et al, 1996a).…”

“…Testing was carried out on 100 Spanish sentences, randomly selected from the text test set, uttered by a four speakers (400 utterances in total). One of the test speakers was included among the 10 training speakers (Vilar et al, 1996a). -EUTRANS-0 and EUTRANS-I: Acoustic models of 26 Spanish monophone units were also used.…”

A Canonical Semi-Deterministic Transducer

“…An important step in the theory of transducers was the development of the algorithm Ostia. Introduced in [29], Ostia was designed for language comprehension tasks [38]. A number of elaborations on the original algorithm have since arisen, many of them aimed at trying to circumvent the restriction to total functions that limited Ostia.…”

Compiling boostexter rules into a finite-state transducer

“…Finite state models have been extensively applied to many aspects of language processing including, speech recognition (Pereira and Riley, 1997), phonology (Kaplan and Kay, 1994), morphology (Koskenniemi, 1984), chunking (Abney, 1991;Bangalore and Joshi, 1999), parsing (Roche, 1999;Oflazer, 1999) and machine translation (Vilar et al, 1999;Bangalore and Riccardi, 2000). Finitestate models are attractive mechanisms for language processing since they (a) provide an efficient data structure for representing weighted ambiguous hypotheses (b) generally effective for decoding (c) associated with a calculus for composing models which allows for straightforward integration of constraints from various levels of speech and language processing.…”