Resolving Translation Ambiguity using Ontological Chains for Cross Language Information Retrieval

1
?????????????????????????Resolving Translation
Ambiguity using Ontological Chains for Cross
Language Information Retrieval

• Institute of Computer and Information Science
• National Chiao Tung University
• Student Je-Wei Liang
• Advisor Dr. Hao-Ren Ke
• Dr. Wei-Pang Yang

2
Outline

• Introduction
• Related Work
• Query Translation
• Word Sense Disambiguation
• Ontology-based CLIR
• Query Translation
• Resolving Translation Ambiguity
• Mono-lingual IR
• Evaluation
• Document set
• Topics
• Relevance Assessment
• Result
• Conclusion
• Reference

3
Introduction What is Cross Language Information
Retrieval?

• Enable users to query in one language and
  retrieve relevant documents in other languages.
• Users neednt know the exact translation of
  their queries.

4
Introduction Why CLIR ?

• Chinese is the most spoken language in the world,
  but web pages are mostly in English.
• Through the aid of CLIR systems, Chinese speakers
  are able to retrieve English documents.

5
Introduction Ambiguity

• Ambiguity occurs when segmentation, translation
  and indexing in CLIR systems.
• Our research focuses on resolving translation
  ambiguity.

6
Motivation and Objectives

• Motivation
• Queries are usually short.
• e.g. ?
• A word may have many possible translations.
• e.g. ?? man, gentleman
• Lack of domain knowledge
• e.g. galleries is related to library
• Objectives
• Design an ontology-based CLIR System
• Query expansion using WordNet.
• Resolving translation ambiguity using ontological
  chains.

7
Related Work
8
Related Work Improved use of Contextual
Information in CLIR Fung98

• ?? and flu have similar context.

• News stories about ??? in Hong Kong.

9
Related Work Improved use of Contextual
Information in CLIR Fung98
10
Related Work WSD using Lexical Chain Barzilay97

• Word meanings are represented by synonym sets
  (synsets)
• Relations defined in WordNet
• Synonym / Antonym
• Hypernym / Hyponym (relation is a kind of)

11
Related Work WSD using Lexical Chain Barzilay97

• A procedure for constructing lexical chains
  follows three steps
• Select a set of candidate words (nouns).
• For each candidate word, find an appropriate
  chain relying on a relatedness criterion among
  members of the chains.
• If it is found, insert the word in the chain and
  update it accordingly.
• Three kinds of relations are defined
• Extra-strongSynonym, 10 point
• StrongHolonym , 7 point
• Medium-strongHypernym , 4 point

12
WSD using Lexical Chain Barzilay97
13
WSD using Lexical Chain Barzilay97

• Machine
• an efficient person
• E.g. the boxer was a magnificent fighting
  machine"

14
WSD using Lexical Chain Barzilay97
Score 11
Score 30
15
Related Work Building a Chinese English WordNet
Chen02
16
Related Work Building a Chinese English WordNet
Chen02
17
Related Work Building a Chinese English WordNet
Chen02
18
Advantages and Disadvantages of Related Work
19
System workflow
20
Query Translation
21
Query Translation

• For each query term, add all its synonyms,
  hypernyms and hyponyms to the query.
• Original query terms are more important than
  newly added terms, so we need to refine term
  weight.
• Term weight in the query can be defined

22
Query Translation An Example

• The query ? is translated to fish, and
  retrieves the following documents by looking up
  WordNet.

23
Query Translation Translation Ambiguity
24
Resolving Translation Ambiguity
25
Ontology Construction

• Each ImageCLEF2004 document belongs to one or
  more categories.
• There are 946 distinct categories.
• Human experts gather related categories to form a
  hierarchy.
• E.g. fish processing and fisherman are
  assigned to the parent node ??.

26
Keyword Extraction

• Ontology Node Representation
• Each node is represented by the term-to-concept
  vector.
• Weight is defined as the product of term
  frequency and inverse concept frequency.
• ltW1,W2, W3, ..,Wngt
• Pick the most important k terms as keywords.

27
Keyword Extraction Example

• Keywords relevant to universities and
  university libraries

28
Building Ontological Chains

• Build an ontological chain for each query.
• For each query, find the most similar N ontology
  leaf nodes.
• Measuring pairwise semantic distance among the
  selected leaf nodes, well obtain a semantic
  graph.
• Find connected components of the network, and
  pick up the strongest component as our
  ontological chain.
• For each node in the chain, add its sibling nodes
  to the chain.
• Calculate mutual information (MI) according to
  the chain for each English query term.
• Pick up terms having MI gt T, and T is the
  threshold.

29
Building Ontological Chains

• Step 1
• Similarity of query Q and ontology leaf node Li
  is defined as the following
• tij is the number of distinct Chinese query terms
  in document j belonging to Li
• N is the number of documents belonging to Li
• E.g. ????????? is similar to fish processing,
  fishwive

30
Building Ontological Chains

• Step2
• Define the semantic distance between 2 ontology
  leaf nodes as
• K is a constant, and D is the path length between
  the 2 nodes.
• E.g. distance between herring and fish
  processing is K/3.

31
Building Ontological Chains

• Calculate pairwise semantic distance and then
  obtain a semantic graph.

32
Building Ontological Chains
Step3 Employ union-find algorithm to find
connected components, and choose the maximum
weighted one.
33
Building Ontological Chains

• Step 4 add fish markets, fisherman
• Step 5,6 For the term ??, pick up man
  instead of gentleman.

34
Monolingual Information Retrieval System
35
Document Vector Representation

• Each document has 3 kind of features
• Terms Wi,j is defined as tf idf
• Categories Wci,j is defined by boolean weighting
  scheme
• Temporal feature Wti,j is defined by boolean
  weighting scheme
• We use cosine measure as our similarity function

36
Query Vector Representation

• Each feature is multiplied by a weighting factor.
• We define 3 temporal operation before, in, after
• E.g. D1 is published in 1898 D2 published in
  1901, and Q is the operation before 1900

37
Evaluation Dataset Description

• ImageCLEF2004 bilingual ad hoc
• task.
• St Andrews University Library
• photographic collection.
• Photos are primarily historic in
• nature from areas in and around
• Scotland.
• Dataset Overview
• 28133 SGML documents consist
• of text and images.
• 946 categories

38
Evaluation Dataset Description
39
Evaluation Topics

• Topics are based on real search request,
  including query logs, and requests from patrons.

40
Evaluation Metrics Mean Average Precision

• Average Precision
• Average of precision at each relevant document
  retrieved.
• E.g. average precision of the query ?????????.
• Mean Average Precision
• Mean of the individual average precision scores.

41
System Demonstration Retrieval
42
System Demonstration Ontology
43
System Demonstration Monolingual
44
System Demonstration Dictionary-based
45
System Demonstration Ontology-based
46
Evaluation Precision/Recall at Top 100

• Ontology-based CLIR performs better than
  dictionary-lookup CLIR.
• Ontology-based CLIR system reaches 85
  performance of monolingual IR system.
• Without Ontology, CLIR reaches only 42
  performance of monolingual IR system.

47
Evaluation 11-point Precision/Recall

• Ontology-based CLIR system performs better than
  dictionary-lookup CLIR.

48
Evaluation Mean Average Precision

• Ontology-based CLIR system performs better than
  dictionary-lookup CLIR.
• Ontology-based CLIR system reaches 92
  performance of monolingual IR system.
• Without Ontology, CLIR reaches only 81
  performance of monolingual IR system.

49
System Demonstration Feedback
50
System Demonstration Feedback
51
System Demonstration Feedback
52
Evaluation Relevance Feedback

• Pick up M retrieved relevant documents as
  positive examples, and N retrieved non-relevant
  documents as negative examples.

53
Discussion

• With ontological chains, CLIR will perform better
  than monolingual IR.
• Without semantic features, documents are
  retrieved only if they have common terms as the
  query.
• catch, fisherman and salting are related to
  the query man and woman processing fish in the
  ontology.
• Ontological chains use semantic features and
  perform better than keyword matching.
• Our similarity function performs better than
  cosine measure.
• Terms are independent in the vector space model.
• Our similarity function have the same effect as
  and operator.

54
Discussion When the query is specific

• Our approach performs a little worse than
  keyword-matching.
• E.g. 1908??????????
• Our approach may expand too much ontology nodes.

55
Discussion When the query is general

• Our approach performs much better than
  keyword-matching.
• E.g. ?????????
• man, woman, processing are general terms
  and appears in many documents.

56
Conclusion

• Ontology can be employed to represent domain
  knowledge in an CLIR systems.
• The proposed ontological chain approach can be
  used to resolve translation ambiguity.
• The proposed ontological chain approach gains
  better precision than others especially when
  translation candidates are very large.

57
Future Work

• Semantic indexing can be used to resolve
  polysemous words.

58
Reference

• Ballesteros98 L. Ballesteros and W.B. Croft,
  Resolving ambiguity for cross language
  retrieval, Proc. 21st annual international ACM
  SIGIR conference on Research and development in
  information retrieval, pp.64-71, 1998.
• Barzilay97 R. Barzilay and M. Elhadad, Using
  Lexical Chains for Text Summarization, ACL/EACL
  Workshop on Intelligent Scalable Text
  Summarization, 1997.
• Carbonell97 J. Carbonell, Y. Yang, R.
  Frederking, R.D. Brown, Y. Geng, and D. Lee,
  "Translingual Information Retrieval A
  Comparative Evaluation," Proc. Fifteenth
  International Joint Conference on Artificial
  Intelligence Vol 1, pp. 708-715, 1997.
• Chen02 H.H. Chen, C.C. Lin and W.C. Lin,
  Building a Chinese-English wordnet for
  translingual applications, ACM Transactions on
  Asian Language Information Processing vol. 1,
  Issue 2, pp.103-122, 2002.
• CLEF04 Cross Language Evaluation Forum,
  avalible at http//clef.iei.pi.cnr.it2002/2004.ht
  ml
• Frakes92 W.B. Frakes, R. Baeza-Yates,
  Information Retrieval, Data Structures
  Algorithms. Prentice Hall, 1992.
• Fung98 P. Fung, L.Y. Yee, An IR Approach for
  Translating New Words from Nonparallel,
  Comparable Texts,Proc. of the 36th Annual
  Conference of the Association for Computational
  Linguistics, pp. 414-420, 1998.

59
Reference

• Gruber93 T. R. Gruber, A translation approach
  to portable ontologies, Knowledge Acquisition,
  pp. 199-220, 1993
• ImageCLEF04 Cross Language Evaluation Forum,
  avalible at http//ir.shef.ac.uk/imageclef2004/
• Kipfer01 B.A. Kipfer and R. L. Chapman, Roget's
  International Thesaurus. , HarperResource, 2001.
• Larkey03 L.S. Larkey and M.E. Connell,
  Structured Queries, Language Modeling, and
  Relevance Modeling in Cross-Language Information
  Retrieval, Information Processing and Management
  Special Issue on Cross Language Information
  Retrieval, 2003.
• Littman98 M.L. Littman, S.T. Dumais, and T.K.
  Landauer,Automatic cross-language information
  retrieval using latent semantic
  indexing,Cross-Language Information Retrieval,
  pp. 5162, 1998.
• Lu02 W.H. Lu, L.F. Chien and H.L. Lee,
  Translation of web queries using anchor text
  mining, ACM Transactions on Asian Language
  Information Processing ,Vol 1, Issue 2,
  pp.159-172, 2002
• Miller95 G. Miller, "Wordnet A Lexical
  Database for English, Proc. of Communications of
  CACM, 1995.

60
Reference

• Miller99 D.R.H. Miller, T. Leek, R.M. Schwartz,
  A hidden Markov model information retrieval
  system, Proc. of the 22nd annual international
  ACM SIGIR conference on Research and development
  in information, pp. 214-221, 1999.
• Nie99 J.Y. Nie, M. Simard, P. Isabelle and R.
  Durand , Cross-Language Information Retrieval
  Based on Parallel Texts and Automatic Mining of
  Parallel Texts from the Web, Proc. of the 22nd
  annual international ACM SIGIR conference on
  Research and development in information, pp.
  74-81, 1999.
• Porter80 M. F. Porter, An algorithm for suffix
  stripping, Program, Vol. 14, No. 3, pp. 130-137,
  1980
• Rocchio71 J. Rocchio, Relevance Feedback in
  Information Retrieval, Prentice-Hall, Inc.,
  1971.
• Salton83 G. Salton and M. J. McGill,
  Introduction to Modern Information Retrieval ,
  McGraw-Hill, 1983
• Savoy03 J. Savoy ,Cross-language information
  retrieval experiments based on CLEF 2000
  corpora, Information Processing Management
  ,Vol. 39, Issue 1, pp. 75-115, 2003.
• Trajan75 R.E. Tarjan, Efficiency of a Good But
  Not Linear Set Union Algorithm, Journal of the
  ACM, Vol 22, Issue 2, pp. 215-225, 1975.

61
Reference

• Xu01 J. Xu, R. Weischedel, and C. Nguyen,
  Evaluating a probabilistic model for
  cross-lingual information retrieval, Proc. 24th
  annual international ACM SIGIR conference on
  Research and development in information retrieval
  , pp. 105-110, 2001
• Zhang02 Y. Zhang and P. Vines, Improved use of
  Contextual Information in Cross-language
  Information Retrieval, ACDS, 2002.