A Machine Translation Toolchain for Polysynthetic Languages

A Machine Translation Toolchain for Polysynthetic Languages Petr Homola Codesign, s.r.o. Palackého 541 252 29 Dobˇrichovice phomola@codesign.cz Abstract We present a set of free tools for building rule-based machine translation systems for polysynthetic languages. As there are no large corpora for most of the “small” lan- guages, it is often impossible to use sta- tistical methods. There are some free MT tools but very little work has been done on polysynthetic languages. The aim of this project is to provide computational tools for morphological and syntactic process- ing for such languages. 1 Introduction The paper describes a set of tools for natural pro- cessing of polysynthetic languages. There are quite a few definitions of polysynthesis. Baker (1996), for example, defines a ‘polysynthesis pa- rameter’ within the Chomskyan framework. How- ever his definition is quite strict and excludes many languages that are traditionally considered polysynthetic (such as Greenlandic). We use Mat- tissen’s (2006) definition which is closer to the un- derstanding of polysynthesis of most researchers in the field. According to her, a language is polysyn- thetic if it contains “complex, polymorphemic verb forms which allow, within one word unit, for com- ponents in the form of non-root bound morphemes with quite ‘lexical’ meaning or optionally for the concatenation of lexical roots”. Due to typological differences from Western languages, polysynthetic languages are quite a challenge for many theories of formal grammar. Our implementation is based on Lexical Func- tional Grammar (Kaplan and Bresnan, 1982; Bres- nan, 2001). The system consists of a morpho- c⃝ 2012 European Association for Machine Translation. logical analyzer, rule-based parser, transfer mod- ule and morphological generator. As an exam- ple of a polysynthetic word, consider the Aymara sentence qullqinipachänwa which corresponds to a complete English sentence: (1) qullqi-ni-pacha-:n-wa money-POSS-EVID-PAST3→3-FOC “Apparently s/he had a lot of money.” The paper is organized as follows: Section 2 de- scribes how we analyze polysynthetic languages morphologically and syntactically. Section 3 gives an overview of the transfer phase. Finally we offer some conclusions in Section 4. 2 Morphological and Syntactic Analysis 2.1 Lexicon Some polysynthetic languages, such as Aymara, have no closed morphological tagset since a stem can be nominalized and/or verbalized several times by adding various derivational suffixes recur- sively without any theoretical limit. The out- put of the morphological analyzer is a set of f(eature)-structures which contain morphosyntac- tic and lexico-semantic information. For example, the f-structure for the Aymara word form uñjsma “I see/saw you” is defined by the following mor- pholexical annotation: (2) (↑ PRED) = ‘see⟨(↑ SUBJ)(↑ OBJ)⟩’ (↑ TENSE) = pres|simple_past (↑ SUBJ PERSON) = 1 ((↑ SUBJ PRED) = ‘pro’) (↑ OBJ PERSON) = 2 ((↑ OBJ PRED) = ‘pro’) The functional equations in (2) encode the verb’s lemma and valency (in the PRED attribute), Proceedings of the 16th EAMT Conference, 28-30 May 2012, Trento, Italy 65 polypersonal agreement (the person of SUBJ and OBJ) and the fact that both arguments can be dropped (in which case the value of the argument’s PRED attribute is ‘pro’). …

Content truncated to 500 words