Automatically Building a Latin Treebank

Tufts University, Medford MA

Investigators

Abstract

Large collections of syntactically parsed sentences (known as treebanks) have proven to be useful for a number of problems in linguistics and computer science, enabling research into the syntax of a language while also providing the cornerstone for such tasks as automatic translation. Building such treebanks, however, is a time- and labor-intensive task since it requires skilled annotators to tag each sentence by hand. This process can be partially automated with the use of computer programs to assign syntactic structure, but the potential for this kind of automation for Latin is unclear, since its free word order and high level of inflection resist accurate automatic analysis. A high level of manual tagging may still be necessary. This project will conduct a feasibility study to determine the resources required to build a multimillion-word Latin treebank. We will develop a preliminary Latin syntactic parser, use it to automate part of the annotation process for a small group of Classical texts (excerpts from Caesar, Vergil, Cicero and Jerome), and then manually correct the results. The immediate significance of this study is an accurate estimate of the costs involved in creating a treebank of a lesser-studied language, and is the necessary first step in building a truly valuable multimillion-word collection of syntactically parsed Latin sentences.