Richard Wicentowski (2004): Multilingual Noise-Robust Supervised Morphological Analysis using the WordFrame Model

• In this paper the author presents a new supervized method for lemmatization, called WordFrame model.
• This new method is compared to existing End-Of-String method and is proven better in most of the cases.
  • A combination of both methods gives even better results.
• The results are evaulated on 30 different languages with median accuracy 97.5%
• The WordFrame model algorithm trains well on noisy data, therefore it can be used in co-training with unsupervised methods.

Both models described in this paper were ment to decompose the word to some basic parts (not morphemes, but similar).

Decomposition of inflection into

• prefix - concatenation of all prefixes
• primary common substring - the stem
• point of suffixation change - phonologicaly induced letter change on the boundary of stem and suffix
• suffix/ending - concatenation of all suffixes of the word

Decomposition of inflection into

• prefix - concatenation of all prefixes
• point of prefixation change - phonologicaly induced letter change on the boundary of first part of stem and prefix
• secondary common substring - the part of stem before stem vowel change
• vowel change - the vowel change inside the stem
• primary common substring - the part of stem after the vowel change
• point of suffixation change - phonologicaly induced letter change on the boundary of stem and suffix
• suffix/ending - concatenation of all suffixes of the word

• R. Wicentowski, 2002, PhD thesis

• Robustness of the algorithm in noisy conditions
• Evaluation on many different languages

• Doesn't do morphological analysis, only lemmatization
• Experiments done only on verbs
• The paper doesn't say, what option the algorithm selects if there are more possible correct results
• The algorithm only uses features based only on the word itself, it doesn't use context
• With information given in this paper, we wouldn't be able to create a program to review the results

• Does the term point of prefixation mean the same as the term morpheme boundary?
• In section 4 of the paper - experimental results presented were done using 10-fold cross validation
  • On what data the autor did the tuning of the models? Aren't the results ? For example ommiting the case of deletion of vowels in stem?
• in section 4.1, Table 5 - shoudn't the WF model give allvays better results than the EOS model? The division of the word in EOS model is simplified division in WF model. Isn't it?

Written by Martin Kirschner