Inferring Morphotactics from Interlinear Glossed Text: Combining Clustering and Precision Grammars

Abstract

In this paper I present a k-means clustering approach to inferring morphological position classes (morphotactics) from Interlinear Glossed Text (IGT), data collections available for some endangered and low-resource languages. While the experiment is not restricted to low-resource languages, they are meant to be the targeted domain. Specifically my approach is meant to be for field linguists who do not necessarily know how many position classes there are in the language they work with and what the position classes are, but have the expertise to evaluate different hypotheses. It builds on an existing approach (Wax, 2014), but replaces the core heuristic with a clustering algorithm. The results mainly illustrate two points. First, they are largely negative, which shows that the baseline algorithm (summarized in the paper) uses a very predictive feature to determine whether affixes belong to the same position class, namely edge overlap in the affix graph. At the same time, unlike the baseline method that relies entirely on a single feature, kmeans clustering can account for different features and helps discover more morphological phenomena, e.g. circumfixation. I conclude that unsupervised learning algorithms such as k-means clustering can in principle be used for morphotactics inference, though the algorithm should probably weigh certain features more than others. Most importantly, I conclude that clustering is a promising approach for diverse morphotactics and as such it can facilitate linguistic analysis of field languages.

Visualizing Inferred Morphotactic Systems

“…The morphological graph that our system visualizes comes from the MOM morphological inference software (Wax, 2014;Zamaraeva, 2016). MOM outputs a directed acyclic graph specified in DOT (Gansner et al, 1993) where nodes are inflectional classes of words and position classes of affixes, and edges reflect the ordering possibilities of those affixes.…”

“…The grammar can be used to produce annotations for additional unglossed data, as in Zamaraeva et al 2017, because the inferred system generalizes beyond the specific combinations of morphemes observed. 1 2 System overview 2.1 Back-end The morphological graph that our system visualizes comes from the MOM morphological inference software (Wax, 2014;Zamaraeva, 2016). MOM outputs a directed acyclic graph specified in DOT (Gansner et al, 1993) where nodes are inflectional classes of words and position classes of affixes, and edges reflect the ordering possibilities of those affixes.…”

Proceedings of the 2nd Workshop on the Use of Computational Methods in the Study of Endangered Languages

“…The goal of the Matrix-ODIN Morphology or 'MOM' system (Wax, 2014;Zamaraeva, 2016) is to extract, from a corpus of IGT, the information required to create a computational morphological grammar of the regularized forms found in the IGT. Specifically, this information includes: (i) a set of affixes, grouped into position classes; (ii) for each affix, the form of the affix (and eventually, the associated morphosyntactic or morphosemantic features, as indicated by the glosses); (iii) for each position class, the inputs it can take (i.e.…”

Extracting Morphophonology from Small Corpora

“…It annotates each morpheme with a label, or gloss tag. When the amount of data is insufficient, the role of such linguistic knowledge in making generalizations becomes more prominent; see (Wax 2014;Zamaraeva 2016) for approaches to extraction of morphological rules that take this path.…”