Probabilistic ParaMor

The ParaMor algorithm for unsupervised morphology induction, which competed in the 2007 and 2008 Morpho Challenge competitions, does not assign a numeric score to its segmentation decisions. Scoring each character boundary in each word with the likelihood that it falls at a true morpheme boundary would allow ParaMor to adjust the confidence level at which the algorithm proposes segmentations. A sliding threshold on segmentation confidence would, in turn, permit a trade off between precision and recall that could optimize F1 or other metrics of interest. Our submission to Morpho Challenge 2009 enriches ParaMor with segmentation confidences by training an off-the-shelf statistical natural language tagger to mimic ParaMor’s morphological segmentations. For a given word, the tagger’s probabilistic confidence that ParaMor would propose the character, c, as the first character of a new morpheme serves as the numeric score of the candidate morpheme boundary that immediately precedes c. We have trained a ParaMor tagger mimic over a development data set of 500,000 unique Hungarian word types. By adjusting the threshold above which the ParaMor mimic proposes morpheme boundaries, we improve ParaMor’s F1 score for Hungarian by 5.9% absolute, from 41.4% to 47.3%. Moreover, by training a probabilistic tagger to emulate the segmentations of a second unsupervised morphology induction system, Morfessor, we are able to combine ParaMor’s segmentation decisions with Morfessor’s to form a single joint segmentation of each word. Our joint ParaMor-Morfessor tagger mimic enhances F1 performance on our Hungarian development set by a further 3.4% absolute, ultimately achieving an F1 score of 50.7%.