Multiple Sequence Alignment for Morphology Induction.
Tzvetan Tchoukalov, Brian Roark, Christian Monson · CLEF (Working Notes) · 2009
MetaMorph is a novel application of multiple sequence alignment (MSA) to natural language morphology induction. Given a text corpus in any language, we sequentially align a subset of the words of the corpus to form an MSA using a probabilistic scoring scheme. We then segment the MSA to produce output analyses. We used this algorithm to compete in the 2009 Morpho Challenge. The F-measure of the analyses produced by MetaMorph are low for the full development corpus, but high for the corpus subsets used to generate the MSA, even surpassing the F-measure of another system used to aid MSA segmentation. This suggests that MSA is an effective algorithm for unsupervised morphology induction and may yet outperform the state-ofthe-art morphology induction algorithms. Future research directions are discussed.