This thesis reports the creation of SANTI-morf (Sistem Analisis Teks Indonesia – morfologi), a rule-based system that performs morphological annotation for Indonesian. The system has been built across three stages, namely preliminaries, annotation scheme creation (the linguistic aspect of the project), and system implementation (the computational aspect of the project).
The preliminary matters covered include the necessary key concepts in morphology and Natural Language Processing (NLP), as well as a concise description of Indonesian morphology (largely based on the two primary reference grammars of Indonesian, Alwi et al. 1998 and Sneddon et al. 2010, together with work in the linguistic literature on Indonesian morphology (e.g. Kridalaksana 1989; Chaer 2008).
As part of this preliminary stage, I created a testbed corpus for evaluation purposes. The design of the testbed is justified by considering the design of existing evaluation corpora, such as the testbed used by the English Constraint Grammar or EngCG system (Voutilanen 1992), the British National Corpus (BNC) 1994 evaluation data , and the training data used by MorphInd (Larasati et al. 2011), a morphological analyser (MA) for Indonesian. The dataset for this testbed was created by narrowing down an existing very large bit unbalanced collection of texts (drawn from the Leipzig corpora; see Goldhahn et al. 2012). The initial collection was reduced to a corpus composed of nine domains following the domain categorisation of the BNC) . A set of texts from each domain, proportional in size, was extracted and combined to form a testbed that complies with the design cited informed by the prior literature.
The second stage, scheme creation, involved the creation of a new Morphological Annotation Scheme (MAS) for Indonesian, for use in the SANTI-morf system. First, a review of MASs in different languages (Finnish, Turkish, Arabic, Indonesian) as well as the Universal Dependencies MAS identifies the best practices in the field. From these, 15 design principles for the novel MAS were devised. This MAS consists of a morphological tagset, together with comprehensive justification of the morphological analyses used in the system. It achieves full morpheme-level annotation, presenting each morpheme’s orthographic and citation forms in the defined output, accompanied by robust morphological analyses, both formal and functional; to my knowledge, this is the first MAS of its kind for Indonesian. The MAS’s design is based not only on reference grammars of Indonesian and other linguistic sources, but also on the anticipated needs of researchers and other users of texts and corpora annotated using this scheme of analysis. The new MAS aims at
The third stage of the project, implementation, consisted of three parts: a benchmarking evaluation exercise, a survey of frameworks and tools, leading ultimately to the actual implementation and evaluation of SANTI-morf.
MorphInd (Larasati et al. 2012) is the prior state-of-the-art MA for Indonesian. That being the case, I evaluated MorphInd’s performance against the aforementioned testbed, both as just5ification of the need for an improved system, and to serve as a benchmark for SANTI-morf. MorphInd scored 93% on lexical coverage and 89% on tagging accuracy. Next, I surveyed existing MAs frameworks and tools. This survey justifies my choice for the rule-based approach (inspired by Koskenniemi’s 1983 Two Level Morphology, and NooJ (Silberztein 2S003) as respectively the framework and the software tool for SANTI-morf.
After selection of this approach and tool, the language resources that constitute the SANTI-morf system were created. These are, primarily, a number of lexicons and sets of analysis rules, as well as necessary NooJ system configuration files. SANTI-morf’s 3 lexicon files (in total 86,590 entries) and 15 rule files (in total 659 rules) are organised into four modules, namely the Annotator, the Guesser, the Improver and the Disambiguator. These modules are applied one after another in a pipeline. The Annotator provides initial morpheme-level annotation for Indonesian words by identifying their having been built according to various morphological processes (affixation, reduplication, compounding, and cliticisation). The Guesser ensures that words not covered by the Annotator, because they are not covered by its lexicons, receive best guesses as to the correct analysis from the application of a set of probable but not exceptionless rules. The Improver improves the existing annotation, by adding probable analyses that the Annotator might have missed. Finally, the Disambiguator resolves ambiguities, that is, words for which the earlier elements of the pipeline have generated two or more possible analyses in terms of the morphemes identified or their annotation.
NooJ annotations are saved in a binary file, but for evaluation purposes, plain-text output is required. I thus developed a system for data export using an in-NooJ mapping to and from a modified, exportable expression of the MAS, and wrote a small program to enable re-conversion of the output in plain-text format. For purposes of the evaluation, I created a 10,000 -word gold-standard SANTI-morf manually-annotated dataset. The outcome of the evaluation is that SANTI-morf has 100% coverage (because a best-guess analysis is always provided for unrecognised word forms), and 99% precision and recall for the morphological annotations, with a 1% rate of remaining ambiguity in the final output.
SANTI-morf is thus shown to present a number of advancements over MorphInd, the state-of-the-art MA for Indonesian, exhibiting more robust annotation and better coverage. Other performance indicators, namely the high precision and recall, make SANTI-morf a concrete advance in the field of automated morphological annotation for Indonesian, and in consequence a substantive contribution to the field of Indonesian linguistics overall.