overview
previous page next page
 |
British Computer Society Natural Language Translation Specialist Group URL: http://www.bcs-mt.org.uk/ |
No. 1, April 1995 ISSN: 1358-8346 http://www.bcs-mt.org.uk/mtreview/1/7.htm |
|
CAT2 — A UNIFICATION-BASED MACHINE TRANSLATION SYSTEM by Ruslan Mitkov Abstract CAT2 is a MT system embodying a unification-based formalism, similar to PATR-II (Sharp 1988, Sharp 1991) and software for the development of grammars, lexicons and translation modules. It was developed at the IAI, Saarbrücken, as a sideline implementation to the Eurotra Project, and has been undergoing constant development and evolution since 1987. Experimental versions of numerous languages have been implemented, including English, German, Spanish, French, Portuguese, Italian, Dutch, Russian, Greek, Korean, and Japanese. It is now being used in pre-industrial projects for a number of commercial firms and academic institutions (Sharp and Streiter 1995).
(a) {cat=s} .[ {cat=np}, {cat=vp} ].
The feature bundles may include any number of simple or complex feature descriptions; simple features have atomic values, for example: 'cat=s', whereas complex features have feature bundles as values, for example: 'agr={num=sing,per=3}'. In addition, since the formalism is implemented in Prolog, a value may be a logical variable, bound to another variable with the same name within the rule; instantiation of one of the variables automatically instantiates the other. The implementation also allows for negative and disjunctive features, implemented in SICStus Prolog using the 'when/2' construct for freezing goal evaluations. The second rule type in tree construction is the 'f-rule' for validating the feature content of partial trees. A simple f-rule for ensuring subject-verb agreement might be coded as follows: (b) {} .[ {cat=np}>>{agr=X}, {cat=vp}>>{agr=X} ].
This rule states that, in a tree configuration containing an NP as left daughter and a VP as right daughter, their agreement features must unify. In practice, our grammars make use of a very small number of b-rules, based on X-bar syntax, (extended) head features (Streiter 1994), and lexically-driven tree construction. The f-rules instantiate various universal and language-specific principles and properties, as well as supplying default values to lexical and phrasal constructions. The tree transduction rules employ analogous rule types: t-rules transform tree structures, and tf-rules copy or transform selected features from source to target trees. The rule formats are similar to b- and f-rules, and again unification underlies the rule application. Since the rules for anaphora resolution in our model do not employ t- or tf-rules, they will not be further described here (see Sharp 1994 for a complete description of the formalism). Recently, CAT2 was extended to be able to handle pronominal anaphora (Mitkov, Choi and Sharp 1995). A pre-industrial prototype of CAT2 has been developed for a dictionary of about 10,000 entries (mainly data processing) in the language pair German-English. For other language pairs (for example, French-German/French-English), experimental versions are available. The prototype accepts free input, especially in the German source pair. Since 1987 CAT2 has been used in various universities as a teaching device and for the definition and processing of language analysis, synthesis and translation. Taking over the long tradition of the University in Saarbrücken in 'electronic language research', the IAI is currently carrying out, apart from CAT2, a number of other application-oriented projects sponsored by the LRE and MLAP programmes of the Commission of the European Union, German ministries, and by private industry. Among these are projects using the most advanced techniques in NLP, such as typed feature structures, for example, ALEP, the Advanced Language Engineering Platform). Other projects are using the well tested CAT2 prototype for industrial validations. References
|