In the beginning of the demo, I will make a few comments on the general facilities of the implementation environment, called MAC-ELU. The most important facilities of the system will be shown. The main part of the demo will be dedicated to the presentation of an elaborated implementation of Romanian morphology which, according to our knowledge, has the widest coverage. A nucleus of the Romanian grammar (a partial rule set for NP) will also be shown.
How it works:
The MAC-ELU is an integrated unification-based system aimed at developing reversible linguistic descriptions. It consists of a morphological analyser/generator, a chart parser, a head-driven generator and a transfer module, all of them relying on unification mechanisms for dealing with grammatical constraints. The morphological processor works on a continuation-classes basis, with the usual clustering of morphemes into distinct dictionaries (called continuation classes). Successful transitions from one cluster to the other, corresponding either to analysis of a word-form structure or to the generation of a word-form by concatenation, are constrained by specific restrictions introduced by means of a powerful macro-definitions mechanisms. The implementation of Romanian morphology (and NP analysis/generation) will be exemplified and the structure of the lexical entries will be discussed. Further development plans will be mentioned.
Who developed it:
The ELU system was developed by ISSCO (Rod Johnson, Mike Rosner, Graham Russel, Afzal Balim, Amy Winarske) , running in ALLEGRO-COMMON LISP on SUN machines. The MAC-ELU version of it, was ported in Machintosh Common Lisp running on Macintoshes by Dan Tufis and Octav Popescu. The ported version includes some new facilities, a menu based interface and it was code-optimized in order to ensure a resonable response time for the smaller machines. The Romanian morphology was implemented by Dan Tufis, Octav Popescu, Lidia Diaconu, Calin Diaconu and Ana-Maria Barbu. Most of the NP rule set is due to Lidia Diaconu, Calin Diaconu and Cristian Dumitrescu.
How it can be obtained:
Contact Dan Tufis: email address:tufis@u1.ici.ro
This is an extension of Tomita's GLR parser, with the significant departure from the original algorithm of working with feature-based grammars. Also, the data structuring introduced by Tomita to take care of the conflicting entries in the LR-tables (graph-structured stack, packed shared parse forests) have been enhanced to allow for non-monadic grammar categories (DAGs). To deal with the complexities problems raised by unification extension of the GLR parser we used special data structures (virtual copying vectors). Although not available test data yet, we expect GULiveR to be provide very good response time in real applications.
This parser, was developed in 1992 by me and Octav Popescu. The initial implementation (Golden Common Lisp for PC) was ported this year on Macintoshes (MCL2.0.1) by Stefan Bruda and Mihai Ciocoiu.
This demo, based on an a larger education software, called PAIL, is a nice tutorial system for learning about parsing. It presents two different paradigms: the old ATN procedural approach and the declarative one supported by a parametrized chart-parser. The graphical interface, the steppers, the browsers, animated graphich, demos and on-line documentation made out of this system a very effective educational tool, highly appreciated by students. The larger PAIL system (which includes besides NLP modules, several other interesting Artificial Intelligence systems - theorem proving, rule-based systems, neural networks, back propagation, constraint satisfaction programming, inductive learning, genetic algorithms) was initially implemented at IDSIA-Lugano by Rod Johnson, Mike Rosner, Paolo Cattaneo and Fabio Baj (with contributions from some others) in Allegro Common Lisp on SUN workstations. The system was ported in MCL for Macintoshes by a common team made of Mike Rosner, Paolo Cattaneo from IDSIA and me, Octav Popescu, Stefan Trausan and Adrian Boangiu from Romanian Academy and ICI.
This system is based on a natural language generator (ALLP), developed by Sue Felshin and Stuart Malone of MIT ATHENA's group. ALLP takes as input a highly verbose interlingual representation of a syntactic structure (GB flavoured) and produces natural language text. KRIL allows for much conceptually specified input, and based on the linguistic information already present in the lexicon automatically generates the lengthy structures needed by ALLP. The overhead added by KRIL is less than 10% of the overal generation time (the medium response time is below 1 second for a 6-8 word sentence). The KRIL generator makes fully transparent the linguistic processing to a client application (such as, for instance, an intelligent tutoring system in second language learning).The KRIL interface was implemented by Dan Tufis.