Analogical learner for natural language processing based on structured string-tree correspondence(sstc)and case-based reasoning
Loading...
Date
2009-05
Authors
Huan Ngee, Lim
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Example-Based Machine Translation (EBMT) is using the similar translation
examples which are retrieved from the Bilingual Knowledge Bank (BKB) to translate
an input sentence. The examples (source and target pairs) in the BKB are annotated
based on a flexible annotation schema known as Synchronous Structured String-
Tree Correspondence (S-SSTC).
Indexing approach has been implemented into our current English-Malay
EBMT to ensure fast retrieval of appropriate examples in the BKB for EBMT to
produce well-formed translations. The source and target example pairs in the BKB
are indexed in word and structure level. The structural indexes are classified
according to different types and structures of examples.
Analogy method is introduced to the EBMT system to increase the accuracy
of translation. Using analogy method, we can identify more appropriate BKB
examples for a given input sentence. From the examples, we derive as many
templates as possible using analogy proportion. These templates are more
structurally related to the input sentence compared to the structural indexes return
by the current approach because the structural indexes are picked based on certain
criteria fixed by the researcher.
After the derivation of the templates, we construct its tree representations
using case-based reasoning method. The purpose of constructing tree
representations is to validate the templates which we have derived. Each template
must correspond to its tree representation.
We have made a comparison between analogy method and structural
indexing approach in term of accuracy of translations and the evaluation results
shown that our new approach achieves better results than existing approach.
Description
Keywords
Learner for natural language , Structured string-tree correspondence(sstc , Case-based reasoning