Systematic Comparison Of Cross-Lingual Projection Techniques For Low-Density Nlp Under Strict Resource Constraints

The field of low-density NLP is often approached from an engineering perspective, and evaluations are typically haphazard - considering different architectures, given different languages, and different available resources - without a systematic comparison. The resulting architectures are then tested on the unique corpus and language for which this approach has been designed. This makes it difficult to truly evaluate which approach is truly the best, or which approaches are best for a given language. In this dissertation, several state-of-the-art architectures and approaches to low-density language Part-Of-Speech Tagging are reimplemented; all of these techniques exploit a relationship between a high-density (HD) language and a low-density (LD) language. As a novel contribution, a testbed is created using a representative sample of seven (HD - LD) language pairs, all drawn from the same massively parallel corpus, Europarl, and selected for their particular linguistic features. With this testbed in place, never-before-possible comparisons are conducted, to evaluate which broad approach performs the best for particular language pairs, and investigate whether particular language features should suggest a particular NLP approach. A survey of the field suggested some unexplored approaches with potential to yield better performance, be quicker to implement, and require less intensive linguistic resources. Under strict resource limitations, which are typical for low-density NLP environments, these characteristics are important. The approaches investigated in this dissertation are each a form of language ifier, which modifies an LD-corpus to be more like an HD-corpus, or alternatively, modifies an HD-corpus to be more like an LD-corpus, prior to supervised training. Each relying on relatively few linguistic resources, four variations of language ifier designs have been implemented and evaluated in this dissertation: lexical replacement, affix replacement, cognate replacement, and exemplar replacement. Based on linguistic properties of the languages drawn from the Europarl corpus, various predictions were made of which prior and novel approaches would be most effective for languages with specific linguistic properties, and these predictions were evaluated through systematic evaluations with the testbed of languages. The results of this dissertation serve as guidance for future researchers who must select an appropriate cross-lingual projection approach (and a high-density language from which to project) for a given low-density language. Finally, all the languages drawn from the Europarl corpus are actually HD, but for the sake of the evaluation testbed in this dissertation, certain languages are treated as if they were LD (ignoring any available HD resources). In order to evaluate how various approaches perform on an actual LD language, a case study was conducted in which part-of-speech taggers were implemented for Tajiki, harnessing linguistic resources from a related HD-language, Farsi, using all of the prior and novel approaches investigated in this dissertation. Insights from this case study were documented so that future researchers can gain insight into what their experience might be in implementing NLP tools for an LD language given the strict resource limitations considered in this dissertation