Improving WSD lexical resources using Topic Signatures

1

• Improving WSD lexical resources using Topic
  Signatures
• Alfio Gliozzo.
• gliozzo_at_itc.it
• ITC-irst
• http//tcc.itc.it/research/textec/topics/disambig
  uation/

2
Outline

• The Knowledge Acquisition Bottleneck
• Augmenting WordNet information
• Topic signature definition
• Topic signatures for WordNet synsets
• Methodologies for Topic Signature Development
• Problems in evaluating Topic Signature Quality
• WSD using topic signature
• Relations between topic signatures and WND
• Topic signature disambiguation?

3
The Knowledge Acquisition Bottleneck

• Many NLP supervised systems suffer from the lack
  of widely available semantic tagged corpora
  marquéz,2000
• Typically 1000-2500 occurrences for each word are
  required in order to train an accurate WSD system
  (i.e. accuracy 78 for the lemma interest using
  2476 examples Bruce and Wiebe,1994
• Performaces decrease drammatically using a small
  corpus (i.e. senseval-2 systems, 10 examples per
  word sense, max accuracy 64)
• NG,97 estimated that the manual annotation
  effort necessary to build a complete sense tagged
  English corpus for WSD would be about 16 man years

4
Solutions (proposals)

• Improving ML techniques optimizing the learning
  curve (Boosting (Escudero et al 2000), SVM,
  Feature selection (Moldovan et al. 2001)
• Acquiring learning examples using automatic
  procedures
• Bootstrap (Learn supervised classifiers in a
  small training corpus in order to produce large
  corpora of low quality examples for
  training,i.e.)
• Acquiring sense tagged corpora from WWW (Mihalcea
  and Moldovan,1999)
• Augmenting WordNet information
• Computational theory to explain synonym and
  polisemy phenomena (??????)

5
Acquiring sense tagged corpora from WWW (Mihalcea
and Moldovan,1999)

• Problem Automatically develop a large scale
  training corpus for WSD
• Solution Use WordNet information to make queries
  about word senses concordances in the retrieved
  documents can be used as examples.
• Queries are made using the synset information
  (synonyms and glosses)
• i.e. query for the noun interest
• Sense1 sense of concern AND (interest OR
  involvement)
• Sense4 fixed charge AND interest percentage of
  amount AND interest
• Results 20 lemmas (7 n, 7 v, 3j, 3r), mean
  polisemy 6, 1080 example manually checked,
  accuracy 91

6
Lexical resources

• In order to reduce the learning corpus dimension
  it is possible to make use of lexical databases
  as additional source of information (i.e.
  WordNet)
• The main problem of such a resources is the lack
  of an uniform and objective methodology, so most
  of the information there present is provided by
  the lexicographer intuition
• WordNet has been criticized for its lack of
  relations between topically related senses and
  the proliferation of word senses Magnini and
  Cavaglià,2000

7
Augmenting WordNet information

• Manual Annotation
• i.e. WN-Domains Magnini and Cavaglià,2000
• Acquisition from annotated corpora
• I.e. Corpus Based domain acquisition Magnini et
  al.,2001
• Acquisition from dictionaries
• Acquisition from non annotated corpora and WWW
• Topic Signatures

8
Topic signature definition

• Concepts are linked to lists of topically related
  words
• A topic signature is defined as a family of
  related terms t,lt(w1,s1)(wi,si) ,where t is
  the topic (I.e the target concept, a word sense),
  wi is a word associated with the topic, with
  strength si
• In order to acquire topic signatures a set of
  documents related to the topic is required
• Topic signatures are extracted comparing the
  distribution of terms inside topic-related and
  non topic-related collections of documents

9
Topic signatures for synsets

• The topic is the word sense
• The collection of texts for a sense is the set of
  document (sentences) in which the word sense
  occurs
• The topic signature is a list of topic related
  words (the most representative words for the
  document collection)
• Example topic signature for the lemma waiter
• Waiter, server - a person whose occupation is to
  serve at table (as in a restaurant) restaurant,
  waitress, dinner, lunch, etc.
• Waiter - a person who waits or awaits hospital,
  station, airport, boyfriend, girlfriend, etc.

10
Acquiring collections of documents related to
senses

• Extracted from collections of documents related
  to the senses Agirre et. Al 2000
• Sense tagged corpora (semcor Miller et al. 1993
  , WSJ part of DSONg and Lee,1996)is the
  collection of all the documents(sentences)
  containing the synset
• Query on collections of documents (Altavista) is
  the collection of documents retrieved by a search
  engine making an appropriate query

11
Obtaining sense related documents from WWW

• Queries are constructed using the information in
  the ontology following the methodology of
  (Mihalcea and Moldovan,1999)
• Four procedures
• 1)Use monosemous synonyms
• 2)Use the defining phrases
• 3)Use synonyms with the AND operator and words
  from the defining phrase with the NEAR operator
• 4)Use synonyms and words from the defining
  phrases with the AND operator
• The procedure i is only applied if the procedure
  i-1fails to retrieve any examples (150 documents
  per sense)

12
Calculating TS weights

• Tokenization and lemmatization of the document
  collections
• For each document collection i (representing the
  ith sense of the word) a vector vfi containing
  the words and their frequencies has been
  extracted (wordj,freqi,j)
• For each vfi the ?2 has been evaluated using the
  vectors belonging to the remaining senses as
  contrast set, obtaining vectors of couples
  (wordj,wi,j) where
• Nb mi,j is the product of the frequency of wordj
  in the whole corpus times the dimension (in
  number of tokens) of the corpus for the sense i
  normalized by the whole corpus dimension
  (expected value of freqi,j assuming true the null
  hypothesis)

13
Example the lemma church

• Church1 Christian church, Cristianity
• catholic 59, spirit 43,protestant 42, rector 32
• Church2

14
Evaluating TS quality

• Its a difficult task
• Strategies
• Manual verification
• WSD Agirre et al.,2000
• WordNet domains comparison Gliozzo,2002

15
Using TS in a WSD task

• Topic signature for WN synsets can be used to
  discriminate two senses of the same word in WSD
  task
• Underlying idea TS for senses of a word are the
  list of the most discriminating terms between
  collections of texts for each sense.
• WSD algorithm
• let Sl-n l 1-n ... L0 l1 ln-1 ln be the
  lemmatized sentence, in which L0 is the lemma to
  disambiguate.
• Let s1, s2, sp be the WN synsets for L0
• Let TSj be the list of words in the for sj
• For each j from 1 to p evaluate
• Select the sense sj such as Sj is maximized

16
WordNet domains comparison

• Topic signature for a synset can be used to
  detect the Semantic Domain of the synset
• Algorithm
• Let
  be the topic
  signature for the synset S
• Let DD1, , Dd be a complete set of domains
  (second level of WND)
• Let Dwi be the normalized set of domain for wordi
• For i form i to d evaluate
  
• where
• Select the domain Di such as Di is maximized