Towards the Automatic Creation of a Wordnet from a Term-based Lexical Network

Transcription

1 Towards the Automatic Creation of a Wordnet from a Term-based Lexical Network Hugo Gonçalo Oliveira, Paulo Gomes (hroliv,pgomes)@dei.uc.pt Cognitive & Media Systems Group CISUC, University of Coimbra Uppsala, July 15, 2010 Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

2 Outline 1 Introduction Lexical ontologies Information extraction Issues Research Goals 2 Approach Clustering for synsets Merging thesauri Assigning terms to synsets 3 Experimentation Preparation Wordnet establishment 4 Concluding remarks Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

3 Lexical ontologies Introduction Lexical ontologies Such as Princeton WordNet [Fellbaum 1998] Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

4 Lexical ontologies Introduction Lexical ontologies Such as Princeton WordNet [Fellbaum 1998] Ontology + lexicon [Hirst 2004] Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

5 Lexical ontologies Introduction Lexical ontologies Such as Princeton WordNet [Fellbaum 1998] Ontology + lexicon [Hirst 2004] Knowledge structured on words and their meanings Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

6 Lexical ontologies Introduction Lexical ontologies Such as Princeton WordNet [Fellbaum 1998] Ontology + lexicon [Hirst 2004] Knowledge structured on words and their meanings Cover the whole language Not based on a specific domain Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

7 Lexical ontologies Introduction Lexical ontologies Such as Princeton WordNet [Fellbaum 1998] Ontology + lexicon [Hirst 2004] Knowledge structured on words and their meanings Cover the whole language Not based on a specific domain Construction and maintenance involve time-consuming human effort! Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

8 Introduction Information extraction from text Information extraction From dictionaries: Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

9 Introduction Information extraction from text Information extraction From dictionaries: 1 basketball, noun a game, also known as hoops, played indoors... game HYPERNYM OF basketball basketball SYNONYM OF hoops Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

10 Introduction Information extraction from text Information extraction From dictionaries: 1 basketball, noun a game, also known as hoops, played indoors... game HYPERNYM OF basketball basketball SYNONYM OF hoops 2 basketball, noun the ball used in playing basketball. ball HYPERNYM OF basketball Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

11 Introduction Information extraction from text Information extraction From dictionaries: 1 basketball, noun a game, also known as hoops, played indoors... game HYPERNYM OF basketball basketball SYNONYM OF hoops 2 basketball, noun the ball used in playing basketball. ball HYPERNYM OF basketball From textual corpora: Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

12 Introduction Information extraction from text Information extraction From dictionaries: 1 basketball, noun a game, also known as hoops, played indoors... game HYPERNYM OF basketball basketball SYNONYM OF hoops 2 basketball, noun the ball used in playing basketball. ball HYPERNYM OF basketball From textual corpora:... team sports, such as basketball, rugby... team sport HYPERNYM OF basketball team sport HYPERNYM OF rugby Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

13 Introduction Natural language is ambiguous Issues Term-based networks are impractical for many applications Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

14 Introduction Issues Natural language is ambiguous Term-based networks are impractical for many applications In the previous example: is hoops a team sport? Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

15 Introduction Issues Natural language is ambiguous Term-based networks are impractical for many applications In the previous example: is hoops a team sport? An example extracted from a Portuguese dictionary: ruína SYNONYM OF queda queda SYNONYM OF habilidade habilidade SYNONYM OF ruína?? Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

16 Introduction Issues Natural language is ambiguous Term-based networks are impractical for many applications In the previous example: is hoops a team sport? An example extracted from a Portuguese dictionary: ruína SYNONYM OF queda queda SYNONYM OF habilidade habilidade SYNONYM OF ruína?? queda can either mean aptitude or downfall! Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

17 Onto.PT Introduction Research Goals Automatic construction of a lexical ontology for Portuguese Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

18 Onto.PT Introduction Research Goals Automatic construction of a lexical ontology for Portuguese Extracted from different sources Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

19 Onto.PT Introduction Research Goals Automatic construction of a lexical ontology for Portuguese Extracted from different sources Manually created thesauri Language dictionaries/encyclopedias Corpora Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

20 Onto.PT Introduction Research Goals Automatic construction of a lexical ontology for Portuguese Extracted from different sources Manually created thesauri Language dictionaries/encyclopedias Corpora Modelled after Princeton WordNet Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

21 Onto.PT Introduction Research Goals Automatic construction of a lexical ontology for Portuguese Extracted from different sources Manually created thesauri Language dictionaries/encyclopedias Corpora Modelled after Princeton WordNet Synsets: groups of synonymous words Synset-based relational triples Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

22 Onto.PT Introduction Research Goals Automatic construction of a lexical ontology for Portuguese Extracted from different sources Manually created thesauri Language dictionaries/encyclopedias Corpora Modelled after Princeton WordNet Synsets: groups of synonymous words Synset-based relational triples WSD based on the knowledge already extracted, not on the context. Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

23 Information flow Approach Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

24 Information flow Approach Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

25 Approach Clustering for synsets Synonymy networks tend to have a clustered structure Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

26 Approach Clustering for synsets Synonymy networks tend to have a clustered structure Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

27 Approach Clustering for synsets Synset discovery (inspired by [Gfeller et al. 2005]) 1 Split the original network into sub-networks and calculate the frequency-weighted adjacency matrix F of each sub-network; Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

28 Approach Clustering for synsets Synset discovery (inspired by [Gfeller et al. 2005]) 1 Split the original network into sub-networks and calculate the frequency-weighted adjacency matrix F of each sub-network; 2 F ij = F ij + F ij δ, 0.5 < δ < 0.5; Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

29 Approach Clustering for synsets Synset discovery (inspired by [Gfeller et al. 2005]) 1 Split the original network into sub-networks and calculate the frequency-weighted adjacency matrix F of each sub-network; 2 F ij = F ij + F ij δ, 0.5 < δ < 0.5; 3 Run MCL [van Dongen 2000], with γ = 1.6, over F for 30 times; Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

30 Approach Clustering for synsets Synset discovery (inspired by [Gfeller et al. 2005]) 1 Split the original network into sub-networks and calculate the frequency-weighted adjacency matrix F of each sub-network; 2 F ij = F ij + F ij δ, 0.5 < δ < 0.5; 3 Run MCL [van Dongen 2000], with γ = 1.6, over F for 30 times; 4 Use the (hard) clustering from each run to create P, a matrix with the probabilities of each pair of words in F belonging to the same cluster; Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

31 Approach Clustering for synsets Synset discovery (inspired by [Gfeller et al. 2005]) 1 Split the original network into sub-networks and calculate the frequency-weighted adjacency matrix F of each sub-network; 2 F ij = F ij + F ij δ, 0.5 < δ < 0.5; 3 Run MCL [van Dongen 2000], with γ = 1.6, over F for 30 times; 4 Use the (hard) clustering from each run to create P, a matrix with the probabilities of each pair of words in F belonging to the same cluster; 5 Remove: (a) big clusters, B, if there is a group of clusters C = C 1, C 2,...C n such that B = C 1 C 2... C n ; (b) clusters completely included in other clusters. Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

32 Approach Merging thesauri Merging synsets from different thesaurus For each synset T i T, select B j B with higher c = T i B j /T i B j 1 B 1 = (diva, beldade, beleza, deidade, deusa, divindade) B 2 = (divindade, deidade, deus, nume) 1 Jaccard coefficient Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

33 Approach Merging thesauri Merging synsets from different thesaurus For each synset T i T, select B j B with higher c = T i B j /T i B j 1 B 1 = (diva, beldade, beleza, deidade, deusa, divindade) B 2 = (divindade, deidade, deus, nume) T 1 = (divindade, diva, deusa) 1 Jaccard coefficient Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

34 Approach Merging thesauri Merging synsets from different thesaurus For each synset T i T, select B j B with higher c = T i B j /T i B j 1 B 1 = (diva, beldade, beleza, deidade, deusa, divindade) B 2 = (divindade, deidade, deus, nume) T 1 = (divindade, diva, deusa) c(t1, B 1 ) = 1 3 c(t1, B 2 ) = Jaccard coefficient Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

35 Approach Merging thesauri Merging synsets from different thesaurus For each synset T i T, select B j B with higher c = T i B j /T i B j 1 B 1 = (diva, beldade, beleza, deidade, deusa, divindade) B 2 = (divindade, deidade, deus, nume) T 1 = (divindade, diva, deusa) c(t1, B 1 ) = 1 3 c(t1, B 2 ) = 1 6 N = B 1 T 1 = (diva, beldade, beleza, deidade, deusa, divindade) 1 Jaccard coefficient Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

36 Mapping methods Approach Assigning terms to synsets Input: Thesaurus T, containing synsets Term-based semantic network, N, where each edge has a type R Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

37 Mapping methods Approach Assigning terms to synsets Input: Thesaurus T, containing synsets Term-based semantic network, N, where each edge has a type R Goal: map a R b N to A R B, (A, B) T Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

38 Mapping methods Approach Assigning terms to synsets Input: Thesaurus T, containing synsets Term-based semantic network, N, where each edge has a type R Goal: map a R b N to A R B, (A, B) T Output: semantic network W, whose nodes are synsets, which relate to other synsets by means of semantic relations (wordnet) Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

39 Procedure 1 Approach Assigning terms to synsets Assignment of a (in a R b) to A: 1 Fix b Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

40 Procedure 1 Approach Assigning terms to synsets Assignment of a (in a R b) to A: 1 Fix b 2 S a T : S ai S a, a S ai Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

41 Procedure 1 Approach Assigning terms to synsets Assignment of a (in a R b) to A: 1 Fix b 2 S a T : S ai S a, a S ai a is not in T? create synset A = (a), a A Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

42 Procedure 1 Approach Assigning terms to synsets Assignment of a (in a R b) to A: 1 Fix b 2 S a T : S ai S a, a S ai a is not in T? create synset A = (a), a A 3 For each S ai S a, Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

43 Procedure 1 Approach Assigning terms to synsets Assignment of a (in a R b) to A: 1 Fix b 2 S a T : S ai S a, a S ai a is not in T? create synset A = (a), a A 3 For each S ai S a, p ai = n ai S ai, n ai = number of terms t j S ai : (t j R b) Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

44 Procedure 1 Approach Assigning terms to synsets Assignment of a (in a R b) to A: 1 Fix b 2 S a T : S ai S a, a S ai a is not in T? create synset A = (a), a A 3 For each S ai S a, p ai = n ai S ai, n ai = number of terms t j S ai : (t j R b) S a1 = (a, c, d, e), p a1 = 3 4 S a2 = (a, f, g), p a2 = 2 3 S a3 = (a, h, i, j), p a3 = 1 4 Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

45 Procedure 1 Approach Assigning terms to synsets Assignment of a (in a R b) to A: 1 Fix b 2 S a T : S ai S a, a S ai a is not in T? create synset A = (a), a A 3 For each S ai S a, p ai = n ai S ai, n ai = number of terms t j S ai : (t j R b) S a1 = (a, c, d, e), p a1 = 3 4 S a2 = (a, f, g), p a2 = 2 3 S a3 = (a, h, i, j), p a3 = 1 4 a Sa1 Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

46 Procedure 1 (stage 2) Approach Assigning terms to synsets Only for semi-mapped triples a R B and A R b Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

47 Approach Assigning terms to synsets Procedure 1 (stage 2) Only for semi-mapped triples a R B and A R b Take advantage of established hypernymy links. Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

48 Approach Assigning terms to synsets Procedure 1 (stage 2) Only for semi-mapped triples a R B and A R b Take advantage of established hypernymy links. Assigning b in A R b Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

49 Approach Assigning terms to synsets Procedure 1 (stage 2) examples and additional cleaning If there is C i C with... C i HYPER OF H A R H, b C i If all C i HYPER OF I i A R I i, triples A R I i can be inferred! Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

50 Approach Assigning terms to synsets Procedure 1 (stage 2) examples and additional cleaning If there is C i C with... C i HYPER OF H A R H, b C i If all C i HYPER OF I i A R I i, triples A R I i can be inferred! If H = (dog) I 1 = (cat), I 1 = (mouse) and C i = (mammal): A = (hair) and R = (PART OF ) A = (animal) and R = (HYPER OF ) Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

51 Approach Alternative mapping procedure Assigning terms to synsets 1 M = term-term matrix based on the adjacencies of the lexical network. Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

52 Approach Assigning terms to synsets Alternative mapping procedure 1 M = term-term matrix based on the adjacencies of the lexical network. 2 Collect all the synsets with a, S a T, and all synsets with b, S b T. Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

53 Approach Alternative mapping procedure Assigning terms to synsets 1 M = term-term matrix based on the adjacencies of the lexical network. 2 Collect all the synsets with a, S a T, and all synsets with b, S b T. 3 For each A S a and B S b, with terms A i A and B j B: sim(a, B) = A B cos(a i, B j ) i=1 j=1 A B Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

54 Approach Alternative mapping procedure Assigning terms to synsets 1 M = term-term matrix based on the adjacencies of the lexical network. 2 Collect all the synsets with a, S a T, and all synsets with b, S b T. 3 For each A S a and B S b, with terms A i A and B j B: sim(a, B) = A B cos(a i, B j ) i=1 j=1 A B 4 Select the pair of synsets with the highest similarity Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

55 Experimentation Preparation Resources used (only nouns) PAPEL 2 lexical network Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

56 Experimentation Resources used (only nouns) Preparation PAPEL 2 lexical network Hypernymy, part-of and member-of triples Synonymy instances Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

57 Experimentation Resources used (only nouns) Preparation PAPEL 2 lexical network Hypernymy, part-of and member-of triples Synonymy instances Huge synonymy sub-network with 16k nodes!!! Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

58 Experimentation Resources used (only nouns) Preparation PAPEL 2 lexical network Hypernymy, part-of and member-of triples Synonymy instances Huge synonymy sub-network with 16k nodes!!! TeP 3 thesaurus OpenThesaurus.PT (OT) 4 CLIP = clustered PAPEL TOP = TeP merged with OT, merged with CLIP Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

59 Resulting Thesaurus Experimentation Preparation Words Synsets TeP OT CLIP TOP Quantity 17,158 5,819 23,741 30,554 Ambiguous 5, ,196 13,294 Most ambiguous Quantity 8,254 1,872 7,468 9,960 Avg. size Biggest Table: (Noun) thesauruses in numbers. Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

60 Experimentation Preparation Clustered sub-network of PAPEL example Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

61 Manual validation Experimentation Preparation Sample Correct Incorrect N/A Agreement CLIP 519 sets 65.8% 31.7% 2.5% 76.1% CLIP 310 sets 81.1% 16.9% 2.0% 84.2% TOP 480 sets 83.2% 15.8% 1.0% 82.3% TOP 448 sets 86.8% 12.3% 0.9% 83.0% Table: Results of manual synset validation. CLIP and TOP only consider synsets with 10 or less words. The quality is higher for smaller synsets. Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

62 Resulting WordNet Experimentation Wordnet establishment Hypernym of Part of Member of Term-based triples 62,591 2,805 5,929 Mapped 27,750 1,460 3,962 1st Same synset Already present 3, Semi-mapped triples 7, Mapped nd Could be inferred Already present Synset-based triples 23,572 1,416 3,783 Table: Results of triples mapping Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

63 Automatic validation Experimentation Wordnet establishment For each triple, A R B 1 Compile a set of textual patterns denoting R, e.g.: (hypo) é um uma (tipo forma variedade...)* de (hyper) (whole/group) é um (grupo conjunto...) de (part/member) Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

64 Automatic validation Experimentation Wordnet establishment For each triple, A R B 1 Compile a set of textual patterns denoting R, e.g.: (hypo) é um uma (tipo forma variedade...)* de (hyper) (whole/group) é um (grupo conjunto...) de (part/member) 2 Score the triple with the help of Google: A B found(a i, B j, R) score = i=1 j=1 A B Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

65 Automatic validation Experimentation Wordnet establishment For each triple, A R B 1 Compile a set of textual patterns denoting R, e.g.: (hypo) é um uma (tipo forma variedade...)* de (hyper) (whole/group) é um (grupo conjunto...) de (part/member) 2 Score the triple with the help of Google: A B found(a i, B j, R) score = i=1 j=1 A B Relation Sample size Validation Hypernymy of 419 synsets 44,1% Member of 379 synsets 24,3% Part of 290 synsets 24,8% Table: Automatic validation of triples Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

66 Concluding remarks Concluding remarks Our way to achieve WSD without a context continues... Clustering is a suitable alternative for establishing synsets Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

67 Concluding remarks Concluding remarks Our way to achieve WSD without a context continues... Clustering is a suitable alternative for establishing synsets What about for networks not extracted from dictionaries? Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

68 Concluding remarks Concluding remarks Our way to achieve WSD without a context continues... Clustering is a suitable alternative for establishing synsets What about for networks not extracted from dictionaries? Rules can be defined to map terms in triples to synsets Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

69 Concluding remarks Concluding remarks Our way to achieve WSD without a context continues... Clustering is a suitable alternative for establishing synsets What about for networks not extracted from dictionaries? Rules can be defined to map terms in triples to synsets Though some triples remain unmapped... Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

70 Concluding remarks Concluding remarks Our way to achieve WSD without a context continues... Clustering is a suitable alternative for establishing synsets What about for networks not extracted from dictionaries? Rules can be defined to map terms in triples to synsets Though some triples remain unmapped... Future: Evaluate the alternative mapping method Exploit other resources: e.g. Wiktionary and Wikipedia Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

71 References The end Christiane Fellbaum, editor (1998). WordNet: An Electronic Lexical Database (Language, Speech, and Communication). The MIT Press. Graeme Hirst (2004). Ontology and the lexicon. In Steffen Staab and Rudi Studer, editors, Handbook on Ontologies, International Handbooks on Information Systems, pages Springer. S. M. van Dongen (2000). Graph Clustering by Flow Simulation. Ph.D. thesis, University of Utrecht. David Gfeller, Jean-Cédric Chappelier and Paulo De Los Rios (2005). Synonym Dictionary Improvement through Markov Clustering and Clustering Stability. In Proc. of International Symposium on Applied Stochastic Models and Data Analysis (ASMDA), pages Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24

72 The end Thank you! Gonçalo Oliveira & Gomes (CISUC) TextGraphs-5 Uppsala, July 15, / 24