OBTAINING A VTG TRANSLATION ENSURE THAT THESE MARKED CLAUSES ARE RE...

5. Obtaining a VTG Translation

ensure that these marked clauses are really

missing main verbs. We segment and tag the

We translate VTGs by using verb redundancy

Chinese source sentence using the Stanford

in related documents: if the VTG was translated

Chinese segmenter and the CRF Chinese POS

in other places in related documents, the existing

tagger developed by Purdue University. If we

translations can be reused. Related documents

find a verb phrase in the Chinese source

are likely to use a good translation for a specific

sentence that was not aligned with any English

VTG as it is used in a similar context. A verb’s

words in the SMT alignment tables, then we

aspect and tense can be directly determined by

label it as a verb translation gap (VTG) and

referencing the corresponding MT examples and

confirm that the marking was correct.

their contexts. If, unfortunately, a given VTG

Newswires and blog articles are retrieved for

did not have any other translation record, then

the VTG will not be processed.

five queries which served as our experimental

To do this, our system first builds verb phrase

test bed. The queries are open-ended and on av-

tables from relevant documents and then uses

erage, answers were 30 sentences in length.

the tables to translate the VTG. We use two verb

Q1: Who/What is involved in Saddam Hussein's trial

phrase tables: one is built from a collection of

Q2: Produce a biography of Jacques Rene Chirac Q3: Describe arrests of person from Salafist Group for

MT documents before any query and is called

Preaching and Combat

the “Static Verb Phrase Table”, and the other

Q4: Provide information on Chen Sui Bian Q5: What connections are there between World Cup games and

one is dynamically built from the retrieved

stock markets?

relevant MT documents for each query and is

We used MT documents retrieved by IR for

called the “Dynamic Verb Phrase Table”.

each query to build the Dynamic Verb Phrase

The construction procedure is the same for

Table. We tested the system on 18,886 MT

both. Given a set of related MT documents and

sentences from the retrieved MT documents for

their MT alignments, we collect all Chinese verb

all of the five queries. Among these MT

phrases and their translations along with their

sentences, 1,142 sentences were detected and

frequencies and contexts.

modified (6 % of all retrieved MT sentences).

One key issue is to decide appropriate

contextual features of a verb. A number of