4. Finding the Main Verb Position
and before the response generator selects
sentences for the answer. It modifies any MT
Chinese ordering differs from English mainly
documents retrieved by the embedded
in clause ordering (Wang et al., 2007) and
information retrieval system that are missing a
within the noun phrase. But within a clause
main verb. All MT results are provided by a
centered by a verb, Chinese mostly uses a SVO
phrase-based SMT system.
or SV structure, like English (Yamada and
Post-editing includes three steps: detect a
Knight 2001), and we can assume the local
clause with a missing main verb, determine
alignment centered by a verb between Chinese
which Chinese verb should have been translated,
and English is a linear mapping relation. Under
and find an example sentence in the related
this assumption, the translation of “被捕” in the
documents with an appropriate sentence which
above example should be placed in the position
can be used to modify the sentence in question.
between “Saddam” and “.”. Thus, once we find a
To detect clauses, we first tag the corpus using a
VTG, its translation can be inserted into the
Conditional Random Fields (CRF) POS tagger
corresponding position of the target sentence
and then use manually designed regular
using the alignment.
expressions to identify main clauses of the
This assumes, however, that there is only one
sentence, subordinate clauses (i.e., clauses which
VTG found within a clause. In practice, more
are arguments to a verb) and conjunct clauses in
than one VTG may be found in a clause. If we
a sentence with conjunction. We do not handle
choose one of them, we risk making the wrong
adjunct clauses. Hereafter, we simply refer to all
choice. Instead, we insert the translations of both
of these as “clause”. If a clause does not have
VTGs simultaneously. This strategy could result
any POS tag that can serve as a main verb (VB,
in more than one main verb in a clause, but it is
VBD, VBP, VBZ), it is marked as missing a
more helpful than having no verb at all.
main verb.
MT alignment information is used to further
Bạn đang xem 4. - BÁO CÁO KHOA HỌC WHERE''''S THE VERB CORRECTING MACHINE TRANSLATION DURING QUESTION ANSWERING POT