Minimally Supervised Morphological Analysis by Multimodal Alignment David

Minimally Supervised Morphological Analysis by Multimodal Alignment David Yarowsky and Richard Wicentowski 1

Introduction l l The Algorithm capable of inducing inflectional morphological analyses of regular and highly irregular forms. The Algorithm combines four original alignment models based on: • • Relative corpus frequency. Contextual Similarity. Weighted string similarity. Incrementally retrained inflectional transduction probabilities. 2

Lecture’s Subjects l l Task definition. Required and Optional resources. The Algorithm. Empirical Evaluation. 3

Task Definition Consider this task as three steps: • Estimate a probabilistic alignment between • • inflected forms and root forms. Train a supervised morphological analysis learner on a weighted subset of these aligned pairs. Use the result from step 2 to iteratively refine the alignment in step 1. 4

(Example (POS l Definitions: 5

. Task Definition cont l The target output of step 1: 6

Required and Optional resources l For the given language we need: • A table of the inflectional Part of Speech • l (POS). A list of the canonical suffixes. A large text corpus. 7

Required and Optional resources. cont l A list of the candidate noun, verb and adjective roots (from dictionary), and any rough mechanism for identifying the candidates POS of the remaining vocabulary. (not based on morphological analysis). l A list of the consonants and vowels. 8

Required and Optional resources. cont l A list of common function words. Not essential l A distance/similarity tables generated on previously studied languages. If available 9

The Algorithm l Combines four original alignment models: • Alignment by Frequency Similarity. • Alignment by Context Similarity. • Alignment by Weighted Levenshtein Distance. • Alignment by Morphological Transformation Probabilities. 10

Lemma Alignment by Frequency Similarity l The motivating dilemma: sing ? singed VBD sing ? taked VBD sang VBD 11

Lemma Alignment by Frequency. Similarity cont l This Table is based on relative corpus frequency: 12

Lemma Alignment by Frequency. Similarity cont 13

Lemma Alignment by Frequency. Similarity cont l A problem: the true alignments between inflections are unknown in advance. l A simplifying assumption: the frequency ratios between inflections and roots is not significantly different between regular and irregular morphological processes. 14

Lemma Alignment by Frequency. Similarity cont l Similarity between regular and irregular forms: 15

Lemma Alignment by Frequency. Similarity cont l The expected frequency should also be estimable from the frequency of any of the other inflectional variants. l VBD/VBG and VBD/VBZ could also be used as estimators. 16

Lemma Alignment by Frequency. Similarity cont 17

Lemma Alignment by Context Similarity l l Based on contextual similarity of the candidate form. Computing similarity between vectors of weighted and filtered context features. Clustering inflectional variants of verbs (e. g. sipped, sipping, and sip). 18

Lemma Alignment by Context. Similarity cont l Example: CWsubj(AUX|NEG)*Vkeyword. DET? CW*CWobj Shlomo is eating the apple 19

Lemma Alignment by Weighted Levenshtein Distance l l Consider overall stem edit distance. A cost matrix with initial distance costs: initially set to (0. 5, 0. 6, 1. 0, 0. 98) 20

Lemma Alignment by Morphological Transformation Probabilities The goal is to generalize a mapping function via a generative probabilistic model. 21

Lemma Alignment by Morphological Transformation Probabilities l Result table: 22

Lemma Alignment by Morphological Transformation Probabilities cont. <root>+<stem change>+<suffix> <inflection> unique P(inflection | root, suffix, POS)=P(stemchange | root, suffix, POS) 23

Lemma Alignment by Morphological Transformation Probabilities cont. Example: 24

Lemma Alignment by Model Combination and the Pigeonhole Principle l l l No single model is sufficiently effective on its own. The Frequency, Levenshtein and Context Similarity models retain equal relative weight. The Morphological Transformation Similarity model increases in relative weight. 26

Lemma Alignment by Model Combination and the Pigeonhole Principle l Example: 27

Lemma Alignment by Model Combination and the Pigeonhole Principle cont. l l The final alignment is based on the pigeonhole principle. For a given POS a root shouldn't have more than one inflection nor should multiple inflections in the same POS share the same root. 28

Empirical Evaluation l Performance: 29