Steps taken in Spanish-Basque speech translation using stochastic
erez, M. In´
Electricity and Electronics
University of the Basque Country
Information Systems and Computation
Polytechnic University of Valencia
The goal of this paper is to summarise a work in
progress focused on speech translation making use
of stochastic ﬁnite-state transducers (SFSTs). The
aim of these devices lays on their versatility to inte-
grate acoustic models within translation models.
Our interest lays on Spanish and Basque, ofﬁ-
cial languages in the 600.000 inhabitant Basque Au-
tonomous Community. These two languages, show
remarkable differences in both syntax and morphol-
ogy, as a result they represent a challenge for SFSTs.
In addition, we deal with little linguistic resources
due to the fact that Basque is a minority language.
1 On the use of SFSTs for ST
Finite-state models constitute an elementary frame-
work not only in syntactic pattern recognition but
also in language processing. Particularly, stochas-
tic ﬁnite-state transducers (SFSTs) have proved to
be of use in machine translation of restricted do-
mains. As it is known, SFSTs can be inferred
from positive bilingual samples following GIATI al-
gorithm (Casacuberta and Vidal, 2004; Gonz´
and Casacuberta, 2009). In a few words, given
the bilingual training set, GIATI looks for a mono-
tonic segmentation and next generates a stochas-
tic regular grammar made up of bilingual symbols
(source words together with target phrases), yield-
ing an SFST.
What make SFSTs interesting for speech transla-
tion (ST), besides of the fast decoding algorithms
they rely on, is their versatility to get them integrated
with other ﬁnite-state models.
Decoupled architectures tackle speech translation
in two consecutive decoding steps. Basically, the
ﬁrst step converts speech utterances into text tran-
scription, and the second step consists on text to
text translation of the recognised string. Admit-
tedly, there are different approaches that differ on
the amount and type of information rendered from
the ﬁrst stage to the second one. As integration is
concerned, one of the challenges of speech transla-
tion consists on exploring different ways of integrat-
ing both acoustic and translation knowledge sources
and translation in an attempt to make them collab-
orate. Intuitively, cooperative models might yield
more accurate estimated hypotheses than those mak-
ing decisions in an isolate manner. the most of both
knowledge sources. In (P´
erez et al., 2010) it was
proved that integrated architectures deal with signif-
icantly more accurate hypotheses than decoupled ar-
Apart from the ability to integrate acoustic and
translation models, SFSTs have also proved to al-
low the integration of multiple languages in order to
carry out multi-target speech translation (P´
al., 2007a). As a result, speech was translated si-
multaneously into several languages.
2 Overcoming challenges of Basque
Basque language is a minority language of unknown
origin by contrast to Spanish, which is a Romance
language. While both languages co-exist in the
Basque Autonomous Community, they differ in both
morphology and syntax.
As for morphology, Basque (by contrast to Span-
ish) is very productive in both noun and verbs, with
more than 17 declension cases that can be recur-
sively appended to a lemma. As a result, a Basque
word tends to be translated into Spanish by more
than one word.
Moreover, the morphology of a word might in-
clude syntactic features, e.g. the Basque word
irakasleek means the teachers as the subject of a
transitive clause (Fig. 1).
In order to tackle the rich morphology of Basque,
erez et al., 2007b) phrase-based SFSTs (PB-
SFST) were proposed within GIATI framework.
Those PB-SFSTs represented a step ahead with re-
spect to previous SFSTs since the monotonic bilin-
gual segmentation groups not only words in the tar-
get language but also words in the source language.
In this line, both statistically and linguistically mo-
tivated phrases were explored. Amongst the linguis-
tically motivated phrases morphologically and syn-
tactically motivated ones were distinguished (P´
et al., 2008). Syntactically motivated phrases im-
proved the performance of the system signiﬁcantly.
As a consequence of the rich morphology of
Basque, inﬂected words show a little repetition ra-
tion within the corpus. As an alternative mechanism
to deal with sparsity of data, categorisation was used
yielding a hierarchically arranged SFST in (Justo
et al., 2010) . This approach allowed to categorise
the bilanguage and infer specialised SFSTs for each
category, which, in addition, allowed to integrate the
acoustic models in the same network. While the for-
mulation of the models is neat, experimentally did
not offer much beneﬁts in terms of performance.
Nevertheless, it might had to do with the small di-
mensions of the corpus with which the experiments
were carried out.
Regarding the syntax, Spanish tend to follow SVO
arrangement while Basque would follow SOV ar-
rangement. As a result very long distance align-
ments are frequent, and this is a big deal for SFSTs
under GIATI approach. The SFSTs which we are
dealing with have shown a limited ability to cope
- kas - - le - - k- e -ira-
Figure 1: Analysis of a Basque word.
3 Concluding remarks and further work
Speech translation making use of SFSTs offers a
versatile framework. Recognised utterance and its
translation can be obtained in a single-pass decod-
PB-SFSTs under GIATI approach have shown to
be useful to tackle speech translation between Span-
ish and Basque. PB approach deals with more ac-
curate alignments than word-based one. In addition,
gathering words into phrases helps alignments not
happen at so long distance, and thus overcome one
of the weakness of regular SFSTs. Reordering still
represents an open problem for SFSTs facing this
pair of languages.
Since Basque is a minority language, linguistic
resources are limited. The aforementioned meth-
ods were explored with a restricted domain corpus.
Admittedly, in order to draw solid conclusions we
should experiment with ample-domain corpora. On
this account, we are currently making efforts to col-
lect a EuroParl-like corpus for Basque with text and
F. Casacuberta and E. Vidal. 2004. Machine translation
with inferred stochastic ﬁnite-state transducers. Com-
putational Linguistics, 30(2):205–225.
alez and F. Casacuberta. 2009. GREAT: a
ﬁnite-state machine translation toolkit implementing a
Grammatical Inference Approach for Transducer In-
ference. In EACL workshop on Computational Lin-
guistics Aspects of Grammatical Inference, 24–32
R. Justo, A. P´
erez, M. In´
es Torres, and F. Casacuberta.
2010. Hierarchical ﬁnite-state models for speech
translation using categorization of phrases. In 11th In-
ternational Conference on Intelligent Text Processing
and Computational Linguistics
erez, M. I. Torres, M. T. Gonz´
alez, and F. Ca-
sacuberta. 2007a. An integrated architecture for
speech-input multi-target machine translation. In
Proc. NAACL-HLT, 133–136
erez, M. I. Torres, and F. Casacuberta. 2007b.
Speech translation with phrase based stochastic ﬁnite-
state transducers. In Proc. IEEE ICASSP
erez, M.I. Torres, and F. Casacuberta. 2008. Join-
ing linguistic and statistical methods for Spanish-to-
Basque speech translation. Speech Communication.
erez, M.I. Torres, and F. Casacuberta. 2010. Poten-
tial scope of a fully-integrated architecture for speech
translation. In Proc. EAMT10