Learning Morphosyntactic Analyzers from the Bible via Iterative Annotation Projection across 26 Languages

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1765–1774 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1765 Learning Morphosyntactic Analyzers from the Bible via Iterative Annotation Projection across 26 Languages Garrett Nicolai and David Yarowsky Center for Language and Speech Processing Johns Hopkins University gnicola2@jhu.edu yarowsky@jhu.edu Abstract A large percentage of computational tools are concentrated in a very small subset of the planet’s languages. Compounding the issue, many languages lack the high-quality linguis- tic annotation necessary for the construction of such tools with current machine learning methods. In this paper, we address both is- sues simultaneously: leveraging the high ac- curacy of English taggers and parsers, we project morphological information onto trans- lations of the Bible in 26 varied test languages. Using an iterative discovery, constraint, and training process, we build inflectional lexica in the target languages. Through a combina- tion of iteration, ensembling, and reranking, we see double-digit relative error reductions in lemmatization and morphological analysis over a strong initial system. 1 Introduction The computational processing of languages such as English and Arabic has undeniably benefited from the construction of annotated datasets such as treebanks and morphological databases. Unfor- tunately, the construction of even modestly-sized treebanks is very expensive, requiring hundreds of hours of expert annotation. The construction of computational tools is in turn limited by a lack of supervised training data. One alternative to hand-annotating low- resource languages (LRL) involves using existing tools for a high-resource language (HRL), such as English, and projecting these annotations to the LRL across a parallel corpus. Consider the example in Figure 1: the English sentence is POS- tagged and dependency parsed by tools that have been trained on large amounts of high-quality data. The sentence is word-aligned to its French translation, and the POS tags and dependency relations follow the alignments to annotate the There EX are VBP doubtless RB many JJ different JJ languages NNS in IN the DET world NN . . expl amod amod case det nmod Il EX y a VBP qu' il se rencontre , tant JJ de JJ divers JJ sons NNS dans IN le DT monde NN . . expl amod amod amod case det nmod advmod nsubj nsubj Figure 1: Projection of POS tags and dependency parse from English to French. Black arrows demonstrate left- to-right dependency relations, while red diamonds il- lustrate right-to-left dependency relations. French words. Note that the projection is not lossless: the aligner could not find a French translation of “doubtless”, and has thus been unable to project the RB tag or advmod relation into French. Parallel corpora are rare, and even when they do exist, they often only exist between specific pairs of languages. However, the documentation of a language often begins with the creation of several important documents, including a dictio- nary of key terms, and translations of religious texts. Thus, documents such as the Christian Bible are among the most translated documents in the world (Mayer …

Content truncated to 500 words