|For example, the word
bird maps to Hawk through the is_a relationship. Duck also interlocks
with bird by the same is_a relationship. By sliding along these
relationships, NLPWin uses the knowledge stored in MindNet to identify
the meaning of words in relations to other words.
The Discourse module, pioneered by Simon
Corston- Oliver, takes the data passed up from previous components
and summarizes it. For instance, it can summarize the essence of
a book, similar to Cliff Notes, presenting the key points of the
Meaning Representation Component
At the top of the NLP arch, the Meaning Represent
ation component represents the Holy Grail of computational linguists,
true language understanding. Once in this state, NLPWin has finished
the increasingly abstract parsing of the original text and it stores
the information in MindNet, it is possible to reverse the entire
process to produce meaningful responses.
In other words, the Generation component converts
the abstract, or logical, forms taken directly from NL Text back
into NL Text. By first dissecting and digesting text fed into it
and then synthesizing meaningful responses enables the systemto
engage humans in conversation (dialogue). While many of the previous
attempts at this type of system have
|| focused on narrow vocabularies,
the NLP Group's ambition is to enable broad coverage of entire natural
languages, such as English, Spanish, Japanese, etc.
Applying NLPWin to Machine Translation
Although the research linguists at Microsoft have made
groundbreaking strides in developing the initial components of NLPWin
(with the Word grammar checker perhaps the most notable milestone),
teaching computers to actually understand language remains a distant
goal. Given that the language Generation module appears to depend
on the Meaning Representation component, the successive and cumulative
nature of NLPWin implies that language translation remains beyond
the current capabilities of the system.
Fortunately for the field of Machine
Translation (MT), the NLP Group has found a method to short-circuit
the process. Once it reaches the Logical Form stage, these highly
abstract constructs stored in MindNet it is possible to match or
map to their counterparts in another language. Thus, the system
could perform MT without the machine truly understanding the meaning
of the words.
The creation of the NLPWin Machine Translation
system takes place in two stages: training and runtime.
Figure 3 presents an overview of the MT training
process. The system begins with a pair of equivalent sample sentences
from a database.