paraphrase identification.
This involves sliding around on the lexical similarity dimension to
locate a match (e.g., "canine" against "dog").
Syntactic paraphrasing may also come into play (e.g., matching "Jupiter
has 18 moons" to "Jupiter's 18 moons"). Often both
are required ("How many moons does Jupiter have?" vs. "Jupiter's
18 satellites").
"MindMelding relies on MindNet's path-finding and
lexical similarity routines. Briefly, paths between the least frequent
word in the input graph and other words directly connect to it are
identified. Along these paths, typically, are words that are found
to be similar to one of the endpoints (e.g. looking for paths between
'car' and 'top' might provide paths linked through 'vehicle' or 'hood').
These newly-identified words, which aren't simply similar in meaning
to the original words but, crucially, similar in this particular lexical
context, can now be used for matching if no structures with the original
words can be found. This process is iterated, so that a number of
contextually-similar words can be identified," says Dolan
Although the MindMelding algorithm relies on the type of
graph matching that is intractable (impossible) for the worst cases,
the wide variety of context and linguistic heuristics that the MT
system brings to bear on the matching problem prevents worst case
scenarios from occurring. Nonetheless, carrying out the match efficiently
is still a highly complex challenge.
"We take the Logical Form and try to find
pieces that
match in the stored database mapping [of MindNet] and follow those
to the corresponding link on the English side.
|
|
Grab all those pieces and
sort of Frankenstein
monster-like put them together into a Linked Logical Form. Right now
we are working on using language modeling techniques to smooth out
any differences that make that stitched together Logical Form look
non-native…using statistical techniques, we smooth out any wrinkles
that don't look like what an English Logical Form should look like,"
says Dolan.
Once MindMeld has worked its magic, the
corresponding pieces of target LFs are stitched together to form an
English target LF, which is handed off to the Generation module. "Provided
we've done a good job of assembling a LF, the NLPWin's generation
component reliably maps that LF into a well-formed target-language
sentence," says Dolan. In the example shown in Figure 4, the
English string "Click the highlighted sample text" is generated
from the original Spanish input "Haga clic en el texto de muestra
resaltado."
At runtime, NLP Group's MT system translates all the
English text in the Microsoft Product Support Services Knowledge Base
(KB) into Spanish, allowing users to search the converted KB using
Spanish queries. "As articles are added or updated in English
(which happens a couple thousand times a week), they will be immediately
(re-) translated and posted to the Spanish KB. Occasionally, as the
MT system improves, the entire KB will be retranslated using newer
versions of the MT system. This will happen incrementally, so users
should not experience any down time," notes Richardson. To date,
internal Microsoft studies indicate a high level of satisfaction with
the results obtained using the translated Spanish PSS Knowledge Base. |
