In this paper, the viability of neural network implementations of core technologies (the focus of this paper is on text technologies) for 10 resource-scarce South African languages is evaluated. Neural networks are increasingly being used in place of other machine learning methods for many natural language processing tasks with good results. However, in the South African context, where most languages are resource-scarce, very little research has been done on neural network implementations of core language technologies. In this paper, we address this gap by evaluating neural network implementations of four core technologies for ten South African languages. The technologies we address are part of speech tagging, named entity recognition, compound analysis and lemmatization. Neural architectures that performed well on similar tasks in other settings were implemented for each task and the performance was assessed in comparison with currently used machine learning implementations of each technology. The neural network models evaluated perform better than the baselines for compound analysis, are viable and comparable to the baseline on most languages for POS tagging and NER, and are viable, but not on par with the baseline, for Afrikaans lemmatization.

The development of a hyphenator and compound analyser for Afrikaans The development of two core-technologies for Afrikaans, viz. a hyphenator and a compound analyser is described in this article. As no annotated Afrikaans data existed prior to this project to serve as training data for a machine learning classifier, the core-technologies in question are first developed using a rule-based approach. The rule-based hyphenator and compound analyser are evaluated and the hyphenator obtains an fscore of 90,84%, while the compound analyser only reaches an f-score of 78,20%. Since these results are somewhat disappointing and/or insufficient for practical implementation, it was decided that a machine learning technique (memory-based learning) will be used instead. Training data for each of the two core-technologies is then developed using “TurboAnnotate”, an interface designed to improve the accuracy and speed of manual annotation. The hyphenator developed using machine learning has been trained with 39 943 words and reaches an fscore of 98,11% while the f-score of the compound analyser is 90,57% after being trained with 77 589 annotated words. It is concluded that machine learning (specifically memory-based learning) seems an appropriate approach for developing coretechnologies for Afrikaans.

No abstract

In this paper, we present a project where existing text-based core technologies were ported to Java-based web services from various architectures. These technologies were developed over a period of eight years through various government funded projects for 10 resource-scarce languages spoken in South Africa. We describe the API and a simple web front-end capable of completing various predefined tasks.

Morphological analysis involves investigating the syntactic class of a word but can also extend to the decomposition and syntactic analysis of its underlying morpheme composition. This is especially relevant to languages with an agglutinative writing system where multiple linguistic words are expressed as a single orthographic word. In this paper, we propose a memory-based approach to canonical segmentation using a windowing approach to recover the uncondensed morphemes that differ from the surface form of a word. Additionally, we propose treating the syntactic labelling of morphemes as a sequence labelling task, similar to part of speech tagging. This approach leverages the internal morpheme composition of a word as local context in much the same way that the surrounding sentence of word serves in the disambiguation of its part-of-speech. Both tasks are modelled separately but performed sequentially by cascading the decomposed morphemes of a word into the task of syntactic labelling. When evaluated on four resource-scarce, conjunctively written Nguni languages, the proposed approach achieves an overall accuracy ranging between 82% and 92% which outperforms previously developed rule-based analysers for the same languages.