Poster presented at the Semeval 2015 workshop. Our system clustered words based on their contexts in order to identify their underlying meanings or senses.

1. Duluth : Word Sense Discrimination
in the Service of Lexicography
SemEval 2015 - Task 15
Corpus Pattern Analysis
Ted Pedersen
University of Minnesota, Duluth
tpederse@d.umn.edu
http://senseclusters.sourceforge.net

2. The Task?
Corpus Pattern Analysis
● CPA parsing : syntactic parsing
and semantic role labeling
● CPA clustering: group together
semantically similar contexts
● CPA lexicography: describe verb
patterns based on syntax and
semantics

3. Evaluation Data
● Microcheck (7 verbs, 123-228 instances each):
– appreciate, apprehend, continue, crush,
decline, operate, undertake
● Wingspread (20 verbs, 7-573 instances each):
– adapt, advise, afflict, ascertain, ask, attain,
avert, avoid, begrudge, belch, bludgeon,
bluff, boo, brag, breeze, sue, teeter,
tense, totter, wing

4. Duluth systems
● Participated in Subtask 2
● Viewed as classical word sense discrimination (or
induction) problem
– Given N target words in context, group into
k clusters based on the similarity of the
contexts
● Automatically discovered number of senses
● AKA SenseClusters
– http://senseclusters.sourceforge.net

5. Pre-processing
● Remove non alphanumeric values
● Convert all text to lower case
● Convert all numeric values to a single
generic string

6. 1st
order features
● If each context is represented as a
vector of features, find the
contexts with the most values in
common
● How many words in each context
are the same?
● Contexts with larger number of
shared words are considered to be
clusters

7. 1st
order example
● i operate a machine
● my surgeon will operate on me today
● he can operate the lathe
● your doctor operated with skill and
confidence
● … no matches among the contexts
(other than the target word)

8. 2nd
order co-occurrence features
● If each context is represented as a
vector of features, find the
contexts that have the most
friends in common
● Each (content) word in a context is
replaced by a vector of co-
occurring words

9. 2nd
order co-occurrence example
● Machine → part, drill, shop
● Lathe → part, drill, mill
● Surgeon → scalpel, nurse, prescribe
● Doctor → waiting, nurse, prescribe

10. 2nd
order co-occurrence example
● i operate a (part, drill, shop)
● my (scalpel, nurse, prescribe) will
operate on me today
● he can operate the (part, drill, mill)
● your (waiting, nurse, prescribe)
operated with skill and confidence

11. run1
●
2nd
order co-occurrences
● Features found within contexts
– Words that occur within 8
positions of target verb 2 or
more times
– Target word co-occurrences (tco)
– Stop words retained

12. run2
●
2nd
order co-occurrences
● Features found in WordNet glosses
– Adjacent words that occur 5 or
more times together
– Bigrams (bi)
– Any bigram where both words are
stop word is removed

13. run3
●
1st
order unigrams
● Features found within contexts
– Any non-stop word that occurs 2
or more times in the contexts
– Unigrams (uni)

14. Results
Microcheck Wingspread
run1 .525 .604
run2 .440 .581
run3 .439 .615
baseline .588 .720

15. Results for run1 cluster stopping
N Given Discovered
appreciate 215 2 2
apprehend 123 3 5
continue 203 7 4
crush 170 5 5
decline 201 3 4
operate 140 8 4
undertake 228 2 2
total 1,280 4.3 3.7

16. Lessons?
● Verbs are (still) hard
– Many methods and previous Semeval
tasks geared towards nouns
● External corpus (WordNet) not helpful
● Unigrams surprisingly effective
● Human lexicographer job security is robust
– for now