Persian Morphological Parser using POS Tagger

1
Persian Morphological Parser using POS Tagger

• By
• Ali Azimizadeh (aazimizadeh_at_yahoo.com)
• Mohammad Mehdi Arab (medi.arab_at_gmail.com)
• Center of Speech Technology, SimAva Co.
• Mashhad, Iran
• Presented By
• Arvin Farahmand
• Ryerson University, Dept. of Electrical and
  Computer Engineering.
• Toronto, Canada

2
Outline

• Why a Morphological Parser?
• Persian Morphology
• Problems in Parsing
• Existing Approaches
• Shortcomings
• Our Approach
• Example
• Experimental Results
• Advantages
• Disadvantages

3
Why a morphological parser?

• Morphological parsing is a an important part of
  Natural Language Processing (NLP). This parser
  has been designed as part of the Parsgooyan NLP.
  It is designed to
• Reduce need of a large vocabulary
• Find the syllable of stress in pronunciation
  based on the attached morpheme
• Adjust tags generated by the tagger
• Improve quality of token-to-word converter

4
Persian Morphology

• There are five distinct type of morphemes
• Lexical nouns, adjectives, verbs (mrd, man)
• Grammatical pronouns, prepositions (v, and)
• Derivative particles (napak, unclean)
• Inflective plural suffixes, superlatives
  (ktabha books)
• Clitic possessives (ktabam, my book)

5
Problems in Parsing

• 1. Homographs
• Words that are written the same but are
  pronounced differently. Mainly due to lack of
  written vowels in Persian.
• e.g krdy -gt kardy (you did)
• kordy (Kurdish)

6
Problems in Parsing

• 2. Out Of Vocabulary (OOV) Words
• Most existing NLP systems for the Persian are
  lexicon based.
• What to do for words that are not in the
  vocabulary?

7
Problems in Parsing

• 3. Lack of a tokenization standard
• Morphemes in Persian can be attached or detached.
  They can proceed or precede the word. This makes
  it difficult to correctly tokenize a word as a
  whole.
• e.g. my_rvm -gt my can be may (wine), or mee
  (imperfective marker)

8
Existing Approaches

• The two major approaches, in Persian, thus far
  have been
• Perslex Riazati 1997, which is based on Englex.
  It is a two step morphological parser
• Shiraz NLP (Megerdoomian, 2000), based on feature
  structures and unification

9
Shortcomings

• Although existing approaches give results with
  high accuracies, there are two weaknesses which
  our approach tries to address
• Inability to deal with words that dont exist in
  their vocabulary in a flexible manner
• Inability to deal with homographs

10
Our Approach

• The parser presented uses a rule-based system and
  statistical features of POS.
• This addresses some of the short comings of
  existing approaches.

11
Lexical Morphemes

• In our approach we divide morphemes in Persian
  into two categories
• - First Order Morphemes Morphemes that do not
  have homographs e.g. tan (yours)
• - Second Order Morphemes Morphemes that do have
  homographs e.g. my -gt may (wine), or mee
  (imperfect tense marker)

12
Steps

• Mapping Persian letters are converted to English
  for usage in Festival software
• Tokenization words are separated based on spaces
  and punctuations
• POS Tagger
• first order morphemes are attached to the roots
  without any special rules
• second order morphemes are attached to the
  roots using CART.

13
CART

• CART is a statistical approach that is used here
  in predicting both attached and free morphemes.
• At each branch a yes/no question is answered and
  a final decision is available at the leaf.
• Ours was constructed from 30,000 sets of three
  word groups
• ltprevious word, morpheme, next wordgt

14
Example

• bh mdrsh myrvm (I go to school).
• Step 1 has already been performed and letters are
  mapped to English.
• Step 2 - Tokenizer Output
• bh
• mdrsh
• myrvm

15
Example

• bh mdrsh myrvm (I go to school).
• Step 3 - POS Tagger
• bh ltPREPgt
• mdrsh ltN_SINGgt
• myrvm ???

16
Example

• bh mdrsh myrvm (I go to school).
• My (mee) is the problematic part, but by feeding
  it to CART and making use of previous word mdrsh
  and next word rvm, it is found to be a morpheme.
• myrvm is tagged as ltV_IMPgt

17
Algorithm
18
Training Data

• Training data set used was made up of 2,500,000
  words, divided into 38 tags.
• Almost 99 of all verbs in usage were included in
  the training data.
• Most OOV words encountered were nouns and
  adjectives.

19
Experimental Results
20
Sources of Error

• A large portion of errors were due to proper
  names, specifically those that end in y.
• These were often mistaken with adjectives which
  also often end in y.
• e.g. Tanavoly

21
Advantages - Homographs

• POS Tags are used in resolving the homograph
  problem.
• e.g. mn -gt man (me), or man (three kilos). One is
  a pronoun while the other is a noun.

22
Advantages OOV Words

• Making use of complete POS strengthens the system
  against Out Of Vocabulary (OOV) words.
• e.g. imantan ra hfz knid (protect your faith)
• imantan is tagged N_SING, if tan is separated and
  iman is searched in vocabulary but not found the
  whole word is decided to be a noun. tan is still
  separated.

23
Disadvantages - Homographs

• Homographic separation is still not perfect.
  Homographs that have the same POS can not be
  distinguished.
• E.g. bvd -gt bood (was) bovad (will be)

24
Disadvantages Error Propagation

• If the POS tagger makes a tagging error in one of
  the first words in a sentence the error will
  propagate through the rest of the words in the
  sentence.
• Attached morphemes can be incorrectly tagged as
  words and be made separate from their roots.
• CART needs previous words to be accurately tagged
  for it to work correctly.

25
The End