1. Coarse-grained Word Sense Disambiguation
Jinying Chen, Martha Palmer
March 25th, 2003

2. Outline Motivation Supervised Verb Sense Grouping
Unsupervised Verb Frameset Tagging
Future Work

3. Motivation
Fine-grained WSD are difficult for both human and machine and well-defined sense groups can alleviate this problem (Martha, Hoa, Christiane, 2002)
Potential application in Machine Translation
When building up a WSD corpus, the sense hierarchy can help annotators in sense tagging speed and accuracy (hopefully ? )

4. Outline Motivation Supervised Verb Sense Grouping
What’s VSG?
Using Semantic Features for VSG
Building Decision Tree for VSG
Experiment Results
Unsupervised Verb Frameset Tagging
Future Work

5. What’s VSG? verb sense group verb sense Frameset2 Frameset1
WN1 WN WN3 WN4
WN6 WN7 WN WN5 WN 9 WN10
WN11 WN12 WN WN 14
WN WN20
verb sense

6. Using Semantic Features for VSG
PropBank
Each verb is defined by several framesets
All verb instances belonging to the same frameset share a common set of roles
Roles can be argn (n=0,1,…) and argM-f
Frameset is consistent with Verb Sense Group
Frameset tags and roles are semantic features for VSG

7. Building Decision Tree for VSG
Use c5.0 of DT
3 Feature Sets:
SF (Simple Feature set) works best:
VOICE: PAS, ACT
FRAMESET: 01,02, …
ARGn (n=0,1,2 …) : 0(not occur), 1(occur)
CoreFrame: 01-ARG0-ARG1, 02-ARG0-ARG2,…
ARGM: 0(has not ARGM), 1(has ARGM)
ARGM-f(f=DIS, ADV, …): i (occur i times)

8. Experiment Results
Table 2 Error rate of Decision Tree on five verbs

9. Discussion Simple feature set and simple DT algorithms works well
Potential sparse data problem
Complicate DT algorithms (e.g. with boosting) tend to overfit the data
Complex features are not utilized by the model
Solution: use large corpus, e.g. parsed BNC corpus without frameset annotation

10. Outline Task Description Methodology
Unsupervised Verb Frameset Tagging
EM Clustering for Frameset Tagging
Features
Preliminary Experiment Results
Future Work

11. EM Clustering for Frameset Tagging
we treat a set of features extracted from the parsed sentences as observed variables and assume they are independent given a hidden variable, c:
(1) f1
cluster c
f2 …… fm

12. In the expectation step, we compute the probability of c conditioned on the set of observed features:
(2)
In the maximization step, we re-compute and by maximizing the log-likelihood of all of the observed data.
Repeat the Expectation and Maximization steps for a fixed number of rounds or until the change of the probability parameters and is under a threshold.

13. To do clustering, we compute for each verb instance with the same formula as in (2) and assign this instance to the cluster that has the maximal
To evaluate we count the majority of the instances in a single cluster which have the same gold-standard Frameset. Other instances not in the majority of a cluster are treated as misclassified.

14. Features WordNet classes for Subject: Person, Animate, State, Event, …
WordNet classes for Object
Passivization: 0, 1
Transitivity: 0, 1
PP as adjuncts: location, direction, beneficiary …
Double objects: 0, 1
Clausal complements: 0, 1

15. Preliminary Experiment Results
Table 3 Accuracy of EM clustering on five verbs

16. Outline Task Description Methodology
Unsupervised Verb Frameset Tagging
Future Work

17. Future Work To improve current model by
Refine Subcategorization Extraction
Use More Features
Example:
a. He has to live with this programming work. (live 02 endure)
b. He lived with his relatives. (live 01 inhabit)
To cluster nouns automatically instead of using WordNet to group nouns

18. Thanks!

19. Table 4 lower bound on Decision Tree error rate

20. Table 5 Error rate of DT with different feature sets

21. Table 6 Accuracy of EM clustering on five verbs

22. What’s VSG?
Aggregate the senses of a verb into several groups according to their similarities
Example: Learn
GROUP 1: WN1, WN3 (acquire a skill) GROUP 2: WN2, WN6 (find out)
SINGLETON: WN4 (be a student) SINGLETON: WN5 (teach)
WordNet Meaning (simplified):
1. acquire or gain knowledge or skills -- ("She learned dancing”)
2. hear, find out -- ("I learned that she has two grown-up children“)
3. memorize, con -- (commit to memory; learn by heart)
4. study, read, take -- (be a student of a certain subject; "She is learning for the bar exam")
5. teach, learn, instruct -- ("I learned them French")
6. determine, find out -- ("I want to learn whether she speaks French“)

23. Groups Senses Portuguese German G1 WN1,WN2 desenvolver entwickeln G2
bilden G3
WN8
ausbilden
WN13 G4
WN5
desenvolver-se
WN10
Bilden sich
Table 7 Portuguese and German translations of develop