|
In this dissertation, a new approach is proposed to extract
lexical knowledge from large text corpora. This approach not
only avoid human cost in compiling knowledge, but also gets rid
of inconsistency introduced by human. It first partially
parses text corpora, and then uses simple rules or existing
knowledge to build more useful lexical knowledge. We have used
this approach to extract noun phrases, acquire predicate-
argument structures and classify adjectives. A new model for
machine translation system is also proposed. This model
integrates the qualitative and the quantitative viewpoints to
machine translation systems. Not only does it reduce the
computation time in processing long sentences, but also it
utilizes the lexical knowledge extracted from large corpora.
The model also consists of three components: analysis, transfer
and synthesis, but they have different mechanisms from the
traditional transfer based translation system. The analysis
module is a shallow parser - a probabilistic chunker. In
addition, to extract the indispensable knowledge from input
sentence for enhancing the translation quality, two knowledge
extractors are proposed. The first extractor is used to
acquire noun phrases in input sentences. The second extractor
is to determine the predicate argument structures in the input
sentences. In order to augment the performance of MT system,
some tools are also proposed. A sentence alignment based on
content tags is proved to be very useful in aligning sentence.
Applying such tool will do the good for the acquisition
bilingual lexical knowledge. Besides, a topic identification
algorithm is invented not only for anaphora resolution, but
also for text domain identification.
|