morphological tag of some word

1
Introduction

• morphological tag of some word
• combination of values of morphological categories
  making sense for this word
• morphological ambiguity
• more than one tag is possible for some word forms
• word form own can represent either an adjective
  or a verb
• Its my own car. x I own a car.
• or in Czech e.g. je can be either a verb or a
  pronoun
• Petr je nemocný. x U je vidím.
• morphological analyzer
• returns a morphological tag(s) for a given word
• morphological disambiguation/tagging
• removing the incorrect tags

2
Morphological tagging

• still an open task for highly inflectional
  languages
• statistical methods and/or rule-based techniques
• advantages of the rule-based approach
• transparency
• linguistic interpretability
• manual improvability
• independency
• the manual design of the rules is expensive and
  requires some linguistic knowledge
• attempts to learn the rules automatically using
  ILP and active learning (Nepil, Popelínský 2001
  Nepil et al., 2001)
• manually annotated learning data were still
  needed
• new method of the generating the disambiguation
  rules using only unannotated data.

3
Basic Idea

• homonymy often has a rather accidental nature
• enu the accusative of the fem. ena or 1st sg
  of the verb hnát
• many Czech feminines in the accusative are not
  homonymous with the 1st sg of any verb. And
  conversely...
• a step aside
• grammatical meaning/morphological tag is
  connected with a function of the word in a
  sentence words with the same grammatical
  meaning have the same or similar functions
• words sharing the same function occur in contexts
  which have some common properties
• these properties are manifested in morphological
  tags of the words constituting these contexts

4
Basic Idea

• Back to our example
• all feminines in the accusative occurs in similar
  contexts which differ from contexts of verbs in
  1st sg
• word form enu cannot be both at the same time
• to resolve a particular ambiguity between the
  tags (or sets of tags) X and Y, we have to search
  a corpus for words unambiguously tagged with tags
  X and Y, resp. and find some common properties of
  their contexts. Then we should be able to examine
  context of a word ambiguously tagged with both X
  and Y and determine, which of them can be removed

5
Training Data

• unlimited amount of unanotated data
• we can select only fine learning examples
• quite strong restrictions
• only whole sentences, no numbers, abbreviations,
  interjections, proper names, words unknown to a
  morphological analyzer, no very short sentences,
  sentences without a finite verb, nor sentences
  with an unclear punctuation
• even with these restrictions we get many
  non-grammatical sentences and incorrectly tagged
  words
• learning examples are annotated by the
  morphological analyzer ajka (Sedlácek, 2001),
  each word is labeled with all tags offered by the
  analyzer
• rare readings of some frequent words are removed
  by simple lexical filters

6
Learning Algorithm

• pos examples neg examples domain knowledge
• gt ILP system
• generates (induces) rules covering the positive
  examples without covering the negative ones
• refines a particular rule according to a given
  criterion
• learning examples for resolving X/Y ambiguity
• sentences with some word unambiguously tagged
  X/Y
• learning examples are encoded into Prolog facts
• to resolve a particular ambiguity X/Y we learn
  both two sets of disambiguation rules
• during disambiguation, whenever all rules
  covering certain word fall into the same set, the
  respective tag is retained, and the other
  removed. If both sets contain some rule covering
  the word, the more probable tag is retained.

7
Learning Algorithm

• Construction of the set of rules RS describing
  utmost positive examples and none of the negative
  ones (P and N could be chosen arbitrarily)
• Set the set RS to an empty set
• Set a rule R to a trivial one
• Choose at most P positive examples covered by
  the rule R, but not covered by any rule from
  RS. End if there are no such examples
• Choose at most N negative examples covered by
  the rule R. If there are no such examples, add
  the rule R to the set RS and continue with step 2
• Try to refine (specialize) the rule R to the
  best advantage according to the selected
  positive and negative examples. End if there is
  no posibility of refinement, otherwise continue
  with step 3
• The utility of possible refinements is measured
  by the following formula
• Pcov / Pall Ncov / Nall
• where Pcov (Ncov) stands for the count of
  positive (negative) examples covered by the
  refined rule, Pall (Nall) stands for the count of
  all positive (negative) examples selected in
  steps 2. and 3.
• ILP system INDEED (Nepil 2003) is used for
  refining the rules

8
Experiments

• three experiments have been performed
• the third and the fifth most frequent Czech word
  (se and je) and the subset of the most frequent
  POS ambiguity (words of type vedení) were chosen
• se is either a reflexive pronoun or a vocalized
  form of the preposition s
• je is either a personal pronoun or the 3rd sg of
  the verb být
• vedení type words are some forms of either
  substantives or adjectives
• the evaluation of the generated rules has been
  performed on the manually annotated corpus DESAM
• all occurrences of these three types have been
  used.
• even the badly disambiguated words in
  non-grammatical, but human-parseable sentences
  have been counted as errors caused by the rules

9
Results

• left recall correctly disambiguated / all
  ambiguous words
• right precision correctly disambiguated /
  disambiguated words
• baseline default precision, selection of the
  more probable tag
• frequency portion among all words ambiguous in
  POS
• HMM is Czech HMM-based Tagger (Krbec, Hajic
  2001)
• EXP is Czech Feature-based Tagger (Hajic 2001)
• it should be stated that the comparison with HMM
  and EXP is not quite fair, as they are not
  specialized in solving these three particular
  ambiguities

10
Discussion

• in all cases I had to relax some of the
  principles proposed formerly
• allowing the coverage of a small amount of
  negative examples
• not all unambiguous words can be used, words
  appearing in non-typical contexts have been
  discarded
• je was substituted with words bearing slightly
  different tags
• disadvantages
• difficulty of searching for adequate unambiguous
  substitutes
• it seems that there will be many ambiguities
  unresolvable with our method. e.g., the
  nominative, accusative and vocative cases have
  the same form for all Czech neuters
• on the other hand, the results show that at least
  for some ambiguities quite accurate rules can be
  learned, which could be useful for a partial
  disambiguation or some preprocessing

11
Conclusions and Future Work

• new method of inducing rules for disambiguation
• learning from raw, unannotated data
• promising results, mainly in the accuracy
• limitations constraining a wider/general
  application
• possible improvements
• some kind of (semi?)automatic elimination of
  non-grammatical learning examples should be
  considered
• some rearranging of the tagset
• some rules are very similar to other ones some
  methods of detecting these similarities could be
  useful
• the lexical filters can always be improved
• some improving or lexicalization of the domain
  knowledge?