A Natural Language Generator for Minority Languages

The Bible Translator’s Assistant (TBTA) is a natural language generator (NLG) designed specifically for field linguists doing translation work in minority languages. In particular, TBTA is intended to generate drafts of the narrative portions of the Bible as well as numerous community development articles in a very wide range of languages. TBTA uses the rich interlingua approach. The semantic representations developed for TBTA consist of a controlled English based metalanguage augmented by a feature system designed specifically for minority languages. The grammar in TBTA has two sections: a restructuring grammar and a synthesizing grammar. The restructuring grammar restructures the semantic representations in order to produce a new underlying representation that is appropriate for a particular target language. Then the synthesizing grammar synthesizes the final surface forms. To date TBTA has been tested with four languages: English, Korean, Jula (Cote d’Ivoire) and Kewa (Papua New Guinea). Experiments with the Jula text indicate that TBTA triples the productivity of professional mother tongue translators without any loss of quality. A model of TBTA is shown below in Figure 1. Figure 1. Underlying model of The Bible Translator’s Assistant 1. The Semantic Representations The development of an adequate method of meaning representation for TBTA’s source texts proved to be a challenge. Formal semantics (Cann, 1993; Rosner, 1992), conceptual semantics (Jackendoff, 2002) and generative semantics (Lakoff, 1975) were each considered but found inadequate. Using the foundational principles of Natural Semantic Metalanguage theory, a set of semantically simple English molecules was identified in a principled manner (Wierzbicka, 1996; Goddard, 1998). These semantic molecules serve as the primary lexemes in TBTA’s ontology. The ontology also includes semantically complex lexemes, but each of those lexemes has an associated expansion rule that automatically expands the complex concept in terms of the semantic molecules for those target languages that don’t have a lexicalized semantic equivalent. The feature set developed for TBTA encodes semantic, syntactic and discourse information. Each feature is an exhaustive etic list of the values pertinent to the world’s languages. For example, each nominal is marked for Number, and the possible values are Singular, Dual, Trial, Quadrial and Plural. Each of these values is necessary because some languages morphologically distinguish all five of these categories. Examples of some of the features and their values are listed below in Tables 1 through 4.