Christiane Fellbaum,
Princeton University, USA
Tutorial on Wordnet, 1st Internationl Wordnet Conference, Mysore, India, 
21-25-jan-2002
===============

PRELIMINARY REMARK: 

Many of these comments are based on experiences gained during 
the construction of the Princeton WordNet. WordNet started as an 
experiment testing theories of human semantic memories, and was 
not initially designed as a tool for NLP. WordNet evolved 
gradually and underwent many changes and additions, some of 
which were motivated by specific funding situations. 
The WordNet experiment now permits us the benefit of hindsight; 
many shortcomings of the Princeton WordNet were successfully addressed 
in EuroWordNet. 


DESIGN

Design with a purpose/application(s) in mind 
(e.g. Word Sense Disambiguation, a basic requirement for 
Machine Translation, Information Retrieval, etc. This 
goal would affect the degree of polysemy in the lexicon.)

Observe international standards as proposed in Mysore 
(cf. minutes of business meeting).


METHODOLOGY

1. Use corpora and minimize intuitions, esp. when distinguishing senses
Use test (e.g., Cruse, 1986) to distinguish e.g., hyponyms from synonyms
See also tests listed by Vossen for EWN construction. 

2. Acquisition of contents 

-automatically 
--fast
--requires good Machine Readable Dictionaries 
--requires manual checking

-manually 
--slow, expensive


CONTENTS

1. Which parts of speech should be included? 
Adjectives less important that nouns, verbs. 
Adverbs are "icing on the cake" and may be excluded for lack of money, time

2. Phrasal verbs? Idioms? Noncompositional compound nouns? 

3. Decide on Core Vocabulary
Specific Domains for specific applications

4. Synsets: 

--As there is arguably no absolute synonymy in language, (some) synonym 
sets can have a single member, though synonyms are most useful 
for WSD. But that work can also be done by definitions (glosses) 
and/or example sentences. 

--Glosses should be uniform (e.g., superordinate plus distinguishing 
features)

--Example sentences should be taken from a corpus and/or the web

5. Polysemy: 

Decide what level of fine-grainedness is appropriate for the purpose. 
Ideally, both fine- and coarse-grained sense distinctions should be 
available. This can be achieved by clustering related senses (as was 
done for many verbs in WordNet); the user can have the option of viewing 
either fine or coarse-grained distinctions. 


2. Conceptual-semantic and Lexical Relations:

Number of Relations: 
Find a balance between too many and too few
E.g., how many kinds of meronymy are truly useful? 
Consider distinction between two kinds of hyponymy: ISA and function/role
as in dog ISA mammal; dog HAS ROLE pet

Cross-part of speech links:
--Verb-noun and adjective-noun links show important information 
about subcategorization properties and selectional rectrictions. 
Can serve to disambiguate as well. 

--Morphologically and semantically related words (as currently done for WN 2.0)