The document discusses word sense disambiguation and lexical chain construction using WordNet, focusing on collaborative learning discussions as the corpus. It outlines methods to compute semantic distances and evaluates word-sense pairs to determine their relationships. The results indicate considerable variance in chain lengths and highlight areas for further research, such as utilizing Wikipedia as a resource and improving linguistic tools.

1. WORD SENSE DISAMBIGUATION AND LEXICAL
CHAINS CONSTRUCTION USING WORDNET
The Second Workshop on Natural Language Processing in Support of Learning: Metrics, Feedback and
Connectivity
September 14th 2010, Bucharest,
ROMANIA
Costin-Gabriel Chiru, Andrei Jancă, Traian Rebedea
Politehnica University of Bucharest

2. Summary
 The corpus
 Lexical Chains
 Semantic Distances
 WordNet
 Word sense
disambiguation
 Results
 Further research

3. The corpus
 Chats consisting of 4 or 5 participants,
debating subjects related to collaborative
learning
 High percentage of misspelled words
 Keywords - “wiki” and “forum” – not present
in WordNet with the proper sense
 Utterances not necessarily a correct text in
respect to English grammar

4. Lexical chains
 Lexical chain: set of words where each word has
an acceptable degree of semantic relatedness to
every other word in the set
 Each word must be fitted into a lexical chain –
How?
 The percentage of words in the chain that are
“related” to a word
 Over 90% - word “belongs” to that chain

5. Semantic distance
 When are two words considered to be
“related”?
 Methods of computing a semantic distance
between words – a number in respect to the
strength of the semantic connection
 Most methods use word frequency and the
lowest superordinate => the need for an
ontology

6. WordNet
 Ontology containing information about the
semantic relationships between words
 Words defining the same concept are
grouped into sets - Synsets
 Directed acyclic graph – synsets are nodes,
semantic relationships are vertices =>
consistency with the lower superordinate
concept

7. Semantic distances in WN
 Path length
• If a vertex exists between two nodes, the two
synsets are related through a semantic
relationship
• The length of such a path shows the strength of
a relationship between two senses.
 Conrath-Jiang measure
• Uses the lower superordinate and word
frequency
• Word frequency – number of hits returned by a

8. Word sense disambiguation
 Assigning a sense to an ambiguous word,
based on a context
 Semantic distance is computed for senses,
not for words
 The following scenario might occur : two
words are found to be semantically close, but
not for their right senses

9. Word sense disambiguation (2)
 Context = a window of words
 Window size – trade-off between time and
quality of results
 Our corpus : ideas and the subject of the
utterances often alternate => a large window
size is not necessary

10. Word sense disambiguation (3)
 Each word in the context has a list of senses
 A set of word-sense pairs : for each word in
the window, a sense is assigned
 We must choose the best such set => a score
must be computed for each set
 What evaluation function should compute the
score?

11. Evaluation function for WSD
 The degree of a word-sense pair : The number
of senses related to that sense in a set
 We use WN and semantic distances to
determine if two senses are related
 High degree – higher probability of a correct
word-sense assignment
 The average of all degrees in a set – high
average = high score

12. Evaluation function for WSD (2)
 Problem : many word-sense pairs with very low
degree and few with very high degree
 All degrees should be “packed” around the
average => standard deviation of all degrees.
 Low standard deviation = high score
 Low semantic distances = all senses are closely
related => we need the average and standard
deviation of all distances

13. Results
 Window size : 3-4 utterances
 High threshold for computing semantic
relatedness
 The lowest superordinate is part of the
shortest path between two senses – we
ignore path lengths > 4
 A word is included in a chain if it is related to
90% of the words present in that chain

14. Results (2)
 Vocabulary size for corpus - 1696 words
 With WSD
 Average chain length for entire corpus = 52.68 words
 Number of chains = 649
 Longest chain : 95 distinct words
 Number of unitary chains (chains with a single distinct word) =
394 => 23.21 % of all words are part of such a chain
 Without WSD
 Average chain length for entire corpus = 94.48 words
 Number of chains = 358 , of which 260 are unitary chains.
Therefore, some chains are very long and probably inaccurate
 Longest chain : 756 distinct words (over 40% of the vocabulary
size)

15. Further research
 There is a high dependency between the
linguistic tool (WN is used now) , the corpus and
algorithms for tasks like lexical chaining and
WSD
 Bottlenecks and key points of this system must
be identified
 Wikipedia – can be used as an ontology
(Wikipedia has a category graph), as well as a
relevant corpus
 Implementing a spell-checker to increase the
number of words taken into account

16. Further research (2)
 Each word must be fitted into a lexical chain –
How?
 When is a chain stronger, rather than where
does a word best belong
• “Strength” of a chain = how closely related are
the words
• Output is a set of chains => “strength” of a set of
chains
• State-space searching : the output of the current
lexical chaining algorithm is the initial state, while
the final state is an acceptably strong set of
chains

17. Q & A