Learning Morphological Disambiguation Rules for Turkish

1
Learning Morphological Disambiguation Rules for
Turkish

• Deniz Yuret
• Ferhan Türe
• Koç University, Istanbul

2
Overview

• Turkish morphology
• The morphological disambiguation task
• The Greedy Prepend Algorithm
• Training
• Evaluation

3
Turkish Morphology

• Turkish is an agglutinative language Many
  syntactic phenomena expressed by function words
  and word order in English are expressed by
  morphology in Turkish.
• I will be able to go.
• (go) (able to) (will) (I)
• git ebil ecek im
• Gidebilecegim.

4
Fun with Turkish Morphology
Avrupalilastiramadiklarimizdanmissiniz

• Avrupa Europe
• li European
• las become
• tir make
• ama not able to

• dik we were
• larimiz those that
• dan from
• mis were
• siniz you

5
So how long can words be?

• uyu sleep
• uyut make X sleep
• uyuttur have Y make X sleep
• uyutturt have Z have Y make X sleep
• uyutturttur have W have Z have Y make X sleep
• uyutturtturt have Q have W have Z

6
Morphological Analyzer for Turkish

• masali
• masalNounA3sgPnonAcc ( the story)
• masalNounA3sgP3sgNom ( his story)
• masaNounA3sgPnonNomDBAdjWith ( with
  tables)
• Oflazer, K. (1994). Two-level description of
  Turkish morphology. Literary and Linguistic
  Computing
• Oflazer, K., Hakkani-Tür, D. Z., and Tür, G.
  (1999) Design for a turkish treebank. EACL99
• Kenneth R. Beesley and Lauri Karttunen, Finite
  State Morphology, CSLI Publications, 2003

7
Features, IGs and Tags
masaNounA3sgPnonNomDBAdjWith

• 8 unique tags
• 11084 distinct tags observed in 1M word training
  corpus

• 126 unique features
• 9129 unique IGs

8
Why not just do POS tagging?
from Oflazer (1999)
9
Why not just do POS tagging?

• Inflectional groups can independently act as
  heads or modifiers in syntactic dependencies.
• Full morphological analysis is essential for
  further syntactic analysis.

10
Morphological disambiguation

• Ambiguity rare in English
• lives lives or lifes
• More serious in Turkish
• 42.1 of the tokens ambiguous
• 1.8 parses per token on average
• 3.8 parses for ambiguous tokens

11
Morphological disambiguation

• Task pick correct parse given context
• masalNounA3sgPnonAcc
• masalNounA3sgP3sgNom
• masaNounA3sgPnonNomDBAdjWith
• Uzun masali anlat Tell the long story
• Uzun masali bitti His long story ended
• Uzun masali oda Room with long table

12
Morphological disambiguation

• Task pick correct parse given context
• masalNounA3sgPnonAcc
• masalNounA3sgP3sgNom
• masaNounA3sgPnonNomDBAdjWith
• Key Idea
• Build a separate classifier for each feature.

13
Decision Lists

• If (W çok) and (R1 DA)
• Then W has Det
• If (L1 pek)
• Then W has Det
• If (W AzI)
• Then W does not have Det
• If (W çok)
• Then W does not have Det
• If TRUE
• Then W has Det

• pek çok alanda (R1)
• pek çok insan (R2)
• insan çok daha (R4)

14
Greedy Prepend Algorithm
GPA(data) 1 dlist NIL 2 default-class
Most-Common-Class(data) 3 rule If TRUE Then
default-class 4 while Gain(rule, dlist, data) gt
0 5 do dlist prepend(rule, dlist) 6
rule Max-Gain-Rule(dlist, data) 7 return
dlist
15
Training Data

• 1M words of news material
• Semi automatically disambiguated
• Created 126 separate training sets, one for each
  feature
• Each training set only contains instances which
  have the corresponding feature in at least one of
  their parses

16
Input attributes

• For a five word window
• The exact word string (e.g. WAli'nin)
• The lowercase version (e.g. Wali'nin)
• All suffixes (e.g. Wn, WIn, WnIn, W'nIn,
  etc.)
• Character types (e.g. Ali'nin would be described
  with WUPPER-FIRST, WLOWER-MID, WAPOS-MID,
  WLOWERLAST)
• Average 40 features per instance.

17
Sample decision lists
Acc 0 1 WInI 1 WyI 1 WUPPER0 1 WIzI 1
L1bu 1 Wonu 1 R1mAK 1 Wbeni 0 Wgünü 1
WInlArI 1 Wonlarý 0 WolAyI 0 Wsorunu
(672 rules)
Prop 1 0 WSTFIRST 0 WTürk 1 WSTFIRST
R1UCFIRST 0 L1. 0 WAnAl 1 R1, 0 WyAD 1
WUPPER0 0 WlAD 0 WAK 1 R1UPPER 0 WMilli 1
WSTFIRST R1UPPER0 (3476 rules)
18
Models for individual features
19
Combining models

• masalNounA3sgP3sgNom
• masalNounA3sgPnonAcc
• Decision list results and confidence (only
  distinguishing features necessary)
• P3sg yes (89.53)
• Nom no (93.92)
• Pnon no (95.03)
• Acc yes (89.24)
• score(P3sgNom) 0.8953 x (1 0.9392)
• score(PnonAcc) (1 0.9503) x 0.8924

20
Evaluation

• Test corpus 1000 words, hand tagged
• Accuracy 95.87 (conf. int 94.57-97.08)
• Better than the training data !?

21
Other Experiments

• Retraining on own output 96.03
• Training on unambiguous data 82.57
• Forget disambiguation, lets do tagging with a
  single decision list 91.23, 10000 rules

22
Contributions

• Learning morphological disambiguation rules using
  GPA decision list learner.
• Reducing data sparseness and increase noise
  tolerance using separate models for individual
  output features.
• ECOC, WSD, etc.