Machine Translation for Languages Lacking Bitext via Multilingual Gloss Transduction

Guarani: A case study in resource development for quick ramp­up MT Ahmed Abdelali, James Cowie, Steve Helmreich, Wanying Jin, Maria Pilar Milagros, Bill Ogden, Hamid Mansouri Rad, and Ron Zacharski New Mexico State University Las Cruces, New Mexico 88003 {ahmed, jcowie, shelmrei, wanying, mmilagro, ogden, hamid, raz}@crl.nmsu.edu Abstract In this paper we describe a set of processes for the acquisition of re­ sources for quick ramp­up machine translation (MT) from any language lacking significant machine tracta­ ble resources into English, using the Paraguayan indigenous lan­ guage Guarani as well as Amharic and Chechen, as examples. Our task is to develop a 250,000 mono­ lingual corpus, a 250,000 bilingual parallel corpus, and smaller corpora tagged with part of speech, named entity, and morphological annota­ tions. 1 Introduction In this paper we describe a set of processes for the acquisition of resources for quick ramp­up ma­ chine translation (MT) from any language lacking significant machine tractable resources into Eng­ lish. In previous work (Nirenburg et al. 1998) we developed an elicitation system that guides non­ expert language informants through questions about the ecology, inflectional morphology, and syntax of their language and leads them through a lexicon development task. This information was then used to automatically generate a transfer MT system. Our current approach replaces this sequen­ tial, guided process with the more free­form acqui­ sition of general resources, which could be used by experts to create an MT system. These resources include a 250,000 parallel text corpus, and a 250,000 word monolingual corpus, as well as smaller corpora tagged for part of speech, named entities, and morphological analysis. The collec­ tion methodology varies with the amount of web/electronic resources that can be developed, the availability and location of linguistic experts and native speakers and the political relationships with the countries where the languages are spoken. We illustrate our methodology by describing our ac­ quisition efforts for Amharic, Chechen, and Guarani. For example, one challenge of the Guarani effort is that due to the lack of Guarani/English bilingual speakers it was neces­ sary to use Spanish as an intermediate or "bridge" language. For Chechen, on the other hand, political ramifications had to be overcome. 2 Background The work described here is situated in the need for quick ramp­up machine translation from languages with few machine­tractable resources (online monolingual corpora, bilingual corpora, lexicons, and analyzers) into English. Development of ma­ chine translation systems requires such resources and one could arguably make the case that the quality of the resulting MT system is greatly de­ pendent on the quality and quantity of this lan­ guage data. For example, the quality of MT sys­ tems developed by statistical methods is dependent on the size and quality of the bilingual parallel corpus. This belief is supported by recent experi­ mental work by Banko and Brill (2001). In contrast to earlier studies (for example, Ratnaparkhi 1999 and Henderson 1999) Banko and Brill found that for statistical natural language processing it makes more sense to allocate resources to increase …

Content truncated to 500 words