Browse Prior Art Database

Method for Extracting Translation Pattern Automatically from Examples

IP.com Disclosure Number: IPCOM000105669D
Original Publication Date: 1993-Aug-01
Included in the Prior Art Database: 2005-Mar-20
Document File: 4 page(s) / 104K

Publishing Venue

IBM

Related People

Watanabe, H: AUTHOR

Abstract

Disclosed is a method to extract translation patterns from translation examples automatically.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 34% of the total text.

Method for Extracting Translation Pattern Automatically from Examples

      Disclosed is  a method to extract translation patterns from
translation examples automatically.

      An Example-Based Approach (EBA) has been expoited for Machine
Translation, and this approach uses a large volume of translation
examples or translation patterns, each of which is a pair of parsed
structures of both languages.  However, it is hard to cellect a large
volume of translation patterns.

      Given an input string Ss, and let St be an output string
produced by a translation system TS, Sc be a correct translation of
Ss.  The proposed method is to obtain translation patterns to produce
the correct translation string Sc.  We suppose the following things:

1.  TS is an example-based transfer system (EBTR) which can directly
    use translation patterns.
2.  Parsing can be done with no errors, that is, the parsed structure
    of source and target languagese are correct.
3.  There are a few translation patterns in translation pattern base.

          We can break a problem to obtain translation patterns into
    the following two sub-problems:
    a.  Make Mc which is the correspondeces between Ds and Dc, and
    b.  Find out translation patterns by comparing <Ds,Mt,Dt> with
        <Ds,Mc,Dc>,
where Ds is a parsed structure of the input string, Dt is the output
structure of EBTR, Mt is the correspondences between Ds and Dt linked
by EBTR, and Dc is a parsed structure of Sc.  In the succeeding
sections, (a) and (b) are described.

Mapping - This section describes how to map correspondent items
between Ds and Dc.  The mapping is determined by the following steps:

1.    Perform (2) and (3) for each node Ns in Ds from root to leaves
    in breadth-first order.
2.  Ns is related with a node Nc in Dc such that
    o   Nc is one of translation words of Ns, and
    o   Nc is not related with any Ds nodes other than Ns.
3.  If no related node is found in Dc, and Ns is the root node
        of Ds, then Ns is related with the root of Dc.

Finding Translation Pattern - This section describes how to find
translation patterns.  Suppose that a mapping Mt between Ds and Dt
and a set of translation patterns Pt are given by EBTR as shown in
Fig. 1, and a mapping Mc between Ds and Dc are given as shown in Fig.
2.

               A               A               A               A
           +---+---+       +---+---+       +---+---+       +---+---+
           |       |       |   |   |       |       |       |       |
           B       C       B   X   C       B       C       Y       C
               +---+---+           |           +---+---+   º
+---+---+
               |    ...