Semantic Relatedness for Evaluation of Course Equivalencies

  • 519 views
Uploaded on

 

More in: Technology , Education
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
No Downloads

Views

Total Views
519
On Slideshare
0
From Embeds
0
Number of Embeds
0

Actions

Shares
Downloads
1
Comments
1
Likes
2

Embeds 0

No embeds

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
    No notes for slide

Transcript

  • 1. Introduction Knowledge Sources Related Work First Approach Second Approach Summary References Semantic Relatedness for Evaluation of Course Equivalencies Doctoral Dissertation Defense Beibei Yang Department of Computer Science University of Massachusetts Lowell July 23, 2012
  • 2. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesOutline 1 Introduction 2 Knowledge Sources 3 Related Work 4 First Approach 5 Second Approach 6 Summary
  • 3. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesNLP and Education Many NLP techniques have been adapted to the education field for: automated scoring and evaluation intelligent tutoring learner cognition However, few techniques address the identification of transfer course equivalencies.
  • 4. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesWhy is it important to suggest transfer courseequivalencies? National Association for College Admission Counseling, 2010 “. . . less attention is focused on the transfer admission process, which affects approximately one-third of students beginning at either a four- or two-year institution during the course of their postsecondary careers.” National Center for Education Statistics, 2005 “For students who attained their bachelor’s degrees in 1999–2000, 59.7 percent attended more than one institution during their undergraduate careers and 32.1 percent transferred at least once.”
  • 5. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesUML’s course transfer dictionary
  • 6. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesCourse descriptions C1 : Analysis of Algorithms Discusses basic methods for designing and analyzing efficient algorithms emphasizing methods used in practice. Topics include sorting, searching, dynamic programming, greedy algorithms, advanced data structures, graph algorithms (shortest path, spanning trees, tree traversals), matrix operations, string matching, NP completeness. C2 : Computing III Object-oriented programming. Classes, methods, polymorphism, inheritance. Object-oriented design. C++. UNIX. Ethical and social issues. f : (C1 , C2 ) → n, n ∈ [0, 1] (1) C1 is a course from an external institution. C2 is a course offered at UML. Slide 34
  • 7. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesKnowledge Acquisition Bottleneck Semantic relatedness measures that rely on a traditional knowledge base usually suffer the knowledge acquisition bottleneck. Knowledge acquisition is difficult for an expert system [HRWL83]: Representation mismatch: the difference between the way a human expert states knowledge and the way it is represented in the system. Knowledge inaccuracy: the difficulty for human experts to describe knowledge in terms that are precise, complete, and consistent enough for use in a computer program. Coverage problem: the difficulty of characterizing all of the relevant domain knowledge in a given representation system, even when the expert is able to correctly verbalize the knowledge. Maintenance trap: the time required to maintain a knowledge base.
  • 8. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesSemantic Relatedness Three terms have been used interchangeably in related literature: semantic relatedness, semantic similarity, and semantic distance. Semantic Distance Semantic Relatedness Semantic Similarity Figure : The relations of semantic distance, semantic relatedness, and semantic similarity [BH06].
  • 9. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesSemantic Similarity versus Semantic Relatedness Semantic Similarity animal cat close human cat distant Semantic Relatedness cat paw close cat hand distant
  • 10. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesPopular Knowledge Sources 1 Lexicon-based Resources Dictionaries Thesauri WordNet Cyc 2 Corpus-based Resources Project Gutenberg British National Corpus Penn Treebank 3 Hybrid Resources Wikipedia Wikitionary
  • 11. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesRelated Work on Semantic Relatedness 1 Lexicon-based Dictionary [KF93] Thesaurus [MH91] WordNet [WP94, LC98, HSO98, YP05] 2 Corpus-based Query Expansion [SH06, BMI07, CV07] LSA [LFL98] HAL [BLL98] PMI-IR [Tur01] ESA (Wikipedia) [GM07, GM09] 3 Hybrid Information Content [Res95] Distributional profiling [Moh06, Moh08] Li et al. [LBM03, LMB+ 06] Ponzetto and Strube (Wikipedia) [PS07]
  • 12. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesA Fragment of the WordNet Taxonomy entity.n.01 physical entity.n.01 ❳ ❢❢❢❢❢ ❳❳❳❳❳❳❳❳❳ ❢❢❢❢❢ ❳❳❳❳❳ ❢❢❢❢❢ object.n.01 ❳ matter.n.03 ❳❳ ❢❢❢❢❢ ❳❳❳❳❳❳❳❳❳ ❳❳❳❳❳ ❢❢❢❢❢ ❳❳❳❳❳ ❳❳❳❳❳ ❢❢❢❢❢ ❳❳❳ part.n.02 whole.n.02 solid.n.01 component.n.03 artifact.n.01 crystal.n.01 crystal.n.02 decoration.n.01 gem.n.02 piezoelectric crystal.n.01 adornment.n.01 transparent gem.n.01 jewelry.n.01 ❳ diamond.n.02 ❢❢❢❢❢ ❳❳❳❳❳❳❳❳❳ ❢❢❢❢❢ ❳❳❳❳❳ ❢❢❢❢❢ bracelet.n.02 necklace.n.01
  • 13. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesThe First Approach 1 Semantic relatedness between two concepts: based on their path length and the depth of their common ancestor in the WordNet taxonomy. 2 Semantic relatedness between two words: based on the previous step, and includes POS and WSD. 3 Semantic relatedness between two sentences: constructs two semantic vectors, and takes into account the information content. 4 Word order similarity (optional): “a dog bites a man” & “a man bites a dog” 5 Semantic relatedness between paragraphs 6 Semantic relatedness between courses
  • 14. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesConcept Relatedness Path function: f1 (p) = e−αp (α ∈ [0, 1]) (2) Depth function: eβh − e−βh f2 (h) = (β ∈ [0, 1]) (3) eβh + e−βh Semantic relatedness between concepts c1 and c2 : fword (c1 , c2 ) = f1 (p) · f2 (h) (4)
  • 15. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesSemantic Relatedness Between Words Algorithm 1 Semantic Relatedness Between Words 1: If two words w1 and w2 have different POS, consider them se- mantically distant. Return 0. 2: If w1 and w2 have the same POS and look the same but do not exist in WordNet, consider them semantically close. Return 1. 3: Using either maximum scores or the first sense heuristic to per- form WSD, measure the semantic relatedness between w1 and w2 using Equation 4 . 4: Using the same WSD strategy as the previous step, measure the semantic relatedness between the stemmed w1 and the stemmed w2 using Equation 4 . 5: Return the larger of the two results in steps (3) and (4), i.e., the score of the pair that is semantically closer.
  • 16. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesConstruct a List of Joint Words To measure the semantic relatedness between sentences S1 and S2 , first join them into a unique word set S, with a length of n: S = S1 ∪ S2 = {w1 , w2 , . . . wn }. (5) S1 : introduction to computer programming S2 : introduction to computing environments S: introduction to computer programming computing environments
  • 17. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesConstruct a Lexical Semantic Vector Algorithm 2 Lexical Semantic Vector s1 for S1 ˆ 1: for all words wi ∈ S do 2: if wi ∈ S1 , set sˆ = 1 where sˆ ∈ s1 . 1i 1i ˆ 3: if wi ∈ S1 , the semantic relatedness between wi and each / word w1j ∈ S1 is calculated using algorithm 1 . Set sˆ to the 1i highest score if the score exceeds a preset threshold δ (δ ∈ [0, 1]), otherwise sˆ = 0. 1i 4: Let γ ∈ [1, n] be the maximum number of times a word w1j ∈ S1 is chosen as semantically the closest word of wi . Let the semantic relatedness of wi and w1j be d, and f1j be the number of times that w1j is chosen. If f1j > γ, set sˆ = d/f1j to give a penalty to w1j . This step is called 1i ticketing. 5: end for
  • 18. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesFirst-level Sentence Relatedness TF-IDF: N T F IDF (wi ) = tfi · idfi = tfi · log (6) dfi Semantic vector SV1 for sentence S1 : SV1i = sˆ ·(T F IDF (wi )+ )·(T F IDF (w1j )+ ), 1i (i ∈ [1, n], j ∈ [1, t]) (7)
  • 19. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesFirst-level Sentence Relatedness (1) SV1 · SV2 fsent (S1 , S2 ) = (8) ||SV1 || · ||SV2 ||
  • 20. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesSecond-level Sentence Relatedness Word order similarity: ||Q1 − Q2 || forder (S1 , S2 ) = 1 − (9) ||Q1 + Q2 || Q1 , Q2 : word order vectors of S1 and S2 . Second-level Sentence Relatedness: (2) (1) fsent (S1 , S2 ) = τ ·fsent (S1 , S2 )+(1−τ )·forder (S1 , S2 ), τ ∈ [0, 1] (10)
  • 21. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesSemantic Relatedness Between Paragraphs n m i=1 (maxj=1 fsent (s1i , s2j )) · Ni fpara (P1 , P2 ) = n (11) i=1 Ni Algorithm 3 Semantic Relatedness for Paragraphs 1: If deletion is enabled, given two course descriptions, select the one with fewer sentences as P1 , and the other as P2 . If deletion is disabled, select the first course description as P1 , and the other as P2 . 2: for each sentence s1i ∈ P1 do 3: Calculate the semantic relatedness between sentences using equation 10 for s 1i and each of the sentences in P2 . 4: Find the sentence pair s1i , s2j (s2j ∈ P2 ) that scores the highest. Save the highest score and the total number of words of s1i and s2j . If deletion is enabled, remove sentence s2j from P2 . 5: end for 6: Collect the highest score and the number of words from each run. Use their weighted mean from equation 11 as the semantic relatedness between P1 and P2 .
  • 22. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesSemantic Relatedness Between Courses fcourse (C1 , C2 ) = θ·fsent (T1 , T2 )+(1−θ)·fpara (P1 , P2 ), θ ∈ [0, 1] (12)
  • 23. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesData sets Data Sets MCC Courses UML Courses Total Small 25 24 49 Medium 55 50 105 Large 108 89 197 Table : Number of courses in the data sets
  • 24. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesExperimental Results Compared against the method by Li et al. [LMB+ 06] and TF-IDF [SB88]: Accuracy Comparison Average ranks of the real equivalent courses 100 Enable word order Enable word order Disable word order 20 Disable word order 90 Best case TFIDF TFIDF Li Li 80 15 70 Average rankAccuracy 60 10 50 40 5 30 Best case 20 0 49 105 197 49 105 197 Number of documents Number of documents
  • 25. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesExperimental Results Performance of two word sense disambiguation algorithms: Accuracy Comparison of WSD 100 90 Best case 80 70 Accuracy 60 50 40 30 FIRST SENSE MAX 20 49 105 197 Number of documents
  • 26. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesWhat’s Wrong with WordNet? 91.304 Foundations of Computer Science A survey of the mathematical foundations of Computer Science. Finite automata and regular languages. Stack Acceptors and Context-Free Languages. Turing Machines, recursive and recursively enumerable sets. Decidability. Complexity. This course involves no computer programming. 64 unfiltered words fetched from WordNet acceptor, adjust, arrange, automaton, basis, batch, bent, calculator, car, class, complexity, computer, countable, course, determine, dress, even, finite, fix, foundation, foundation garment, fructify, hardening, imply, initiation, involve, jell, language, linguistic process, lyric, machine, mathematical, naturally, necessitate, numerical, path, place, plant, push-down list, push-down storage, put, recursive, regular, review, rig, run, science, set, set up, sic, sketch, skill, smokestack, specify, speech, stack, stage set, surveil, survey, terminology, turing, typeset, unconstipated, view.
  • 27. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesWhat’s Wrong with WordNet? 91.304 Foundations of Computer Science A survey of the mathematical foundations of Computer Science. Finite automata and regular languages. Stack Acceptors and Context-Free Languages. Turing Machines, recursive and recursively enumerable sets. Decidability. Complexity. This course involves no computer programming. 18 articles fetched from Wikipedia using the second approach Alan Turing, Algorithm, Automata theory, Complexity, Computer, Computer science, Context-free language, Enumeration, Finite set, Finite-state machine, Kolmogorov complexity, Language, Machine, Mathematics, Recursive, Recursive language, Recursively enumerable set, Set theory. Slide 33
  • 28. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesGrowth of Wikipedia and WordNet over the years Growth of English Wikipedia and WordNet 4000000 Articles in Wikipedia 3500000 Synsets in WordNet 3000000 Article/Synset count 2500000 2000000 1500000 1000000 500000 1992 1996 2000 2004 2008 2012 Year
  • 29. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesWordNet versus Wikipedia Fragments of WordNet and Wikipedia Taxonomies WordNet [Root: synset(‘‘technology’’), #depth: 2] # nodes: 25 Wikipedia [Centroid: ‘‘Category:Technology’’, #steps: 2] # nodes: 3583
  • 30. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesExtract a Lexicographical Hierarchy from Wikipedia 1 Let’s assume the knowledge domain is specified, e.g., “Category:Computer science.” 2 Choose its parent as the root, i.e., “Category:Applied sciences.” 3 Use a depth-limited search to recursively traverse each subcategory (including subpages) to build a lexicographical hierarchy with depth D.
  • 31. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesGrowth of the Hierarchy from Wikipedia Depth: 3 Depth: 1 Depth: 2 Total Nodes: 64,407 Total Nodes: 72 Total Nodes: 4,249 Growth of the lexicographical hierarchy constructed from Wikipedia, illustrated in circular trees. A lighter color of the nodes and edges indicates that they are at a deeper depth in the hierarchy.
  • 32. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesLexicographical Hierarchy constructed from Wikipedia Depth (D) Number of concepts at this level 1 71 2 4,177 3 60,158 4 177,955 5 494,039 6 1,848,052 Table : Number of concepts for each depth in the “Category:Applied sciences” hierarchy. The hierarchy only include 1,534,267 distinct articles, out of 5,329,186 articles in Wikipedia. ⇒ Over 71% Wikipedia articles are eliminated.
  • 33. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesGenerate Course Description Features Algorithm 4 Feature Generation (F ) for Course C 1: Tc ← ∅ (clear terms), Ta ← ∅ (ambiguous terms). 2: Generate all possible n-grams (n ∈ [1, 3]) G from C. 3: Fetch the pages whose titles match any of g ∈ G from Wikipedia redirection data. For each page pid of term t, Tc ← Tc ∪ {t : pid}. 4: Fetch the pages whose titles match any of g ∈ G from Wikipedia page title data. If a disambiguation page, include all the terms this page refers to. If a page pid corresponds to a term t that is not ambiguous, Tc ← Tc ∪{t : pid}, else Ta ← Ta ∪ {t : pid}. 5: For each term ta ∈ Ta , find the disambiguation that is on average most related using Equation 4 to the set of clear terms. If a page pid of ta is on average the most related to the terms in Tc , and the relatedness score is above a threshold δ (δ ∈ [0, 1]), set Tc ← Tc ∪ {ta : pid}. If ta and a clear term are different senses of the same term, keep the one that is more related to all the other clear terms. 6: Return clear terms as features. Slide 27
  • 34. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesExample of Course Features C1 : {1134:“Analysis”, 775:“Algorithm”} {41985:“Shortest path problem”, 597584:“Tree traversal”, 455770:“Spanning tree”, 18955875:“Tree”, 1134:“Analysis”, 18568:“List of algorithms”, 56054:“Completeness”, 775:“Algorithm”, 144656:“Sorting”, 8519:“Data structure”, 93545:“Structure”, 8560:“Design”, 18985040:“Data”} C2 : {5213:“Computing”} {21347364:“Unix”, 289862:“Social”, 9258:“Ethics”, 6111038:“Object-oriented design”, 5311:“Computer programming”, 72038:“C++”, 27471338:“Object-oriented programming”, 8560:“Design”} Slide 6
  • 35. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesLexical Semantic Vector An algorithm similar to Algorithm 2 is used to determine each value of an entry of the lexical semantic vector sˆ for features 1i F1 . A semantic vector is defined as: SV1i = sˆ · I(ti ) · I(tj ) 1i (13)
  • 36. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesInformation Content Information content I(t) of a term t: I(t) = γ · Ic (t) + (1 − γ) · Il (t). (14) Category information content Ic (t): log(siblings(t) + 1) Ic (t) = 1 − , (15) log(N ) Linkage information content Il (t): inlinks(pid) outlinks(pid) Il (t) = 1 − · , (16) M AXIN M AXOU T
  • 37. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesDetermine Course Relatedness SV1 · SV2 f (C1 , C2 ) = . (17) ||SV1 || · ||SV2 || f (T1 , T2 ) · (||FT 1 || + ||FT 2 ||) + f (C1 , C2 ) · (||FC1 || + ||FC2 ||) f (course1 , course2 ) = +Ω, ||FT 1 || + ||FT 2 || + ||FC1 || + ||FC2 || (18)
  • 38. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesExperimental Results Randomly select 25 CS courses from 19 universities that can be transferred to UML according to the transfer dictionary. Each transfer course is compared to all 44 CS courses offered at UML. The result is considered correct if the real equivalent course at UML is among the top 3 in the list of highest scores. Algorithm Accuracy Proposed approach 72% Li et al. [LMB+ 06] 52% TF-IDF 32% Table : Accuracy of the second approach against those of Li et al., and TFIDF
  • 39. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesExperimental Results Algorithm Pearson’s correlation p-value TF-IDF 0.730 2 · 10−6 Li et al. [LMB+ 06] 0.570 0.0006 Proposed approach (Features) 0.845 1.13 · 10−9 Proposed approach (Features + IC) 0.851 6.65 · 10−10 Table : Pearson’s correlation of course relatedness scores with human judgments.
  • 40. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesSensitivity Test Testing the Sensitivity of Parameters α, β, and δ 1.0 Pearson Correlation When α Changes (β =0.5, δ =0.2) 0.8 Pearson correlation 0.6 0.4 0.2 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 α 1.0 Pearson Correlation When β Changes (α =0.2, δ =0.2) 0.8 Pearson correlation 0.6 0.4 0.2 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 β 1.0 Pearson Correlation When δ Changes (α =0.2, β =0.5) 0.8 Pearson correlation 0.6 0.4 0.2 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 δ
  • 41. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesSummary Highlight the problem of suggesting transfer course equivalencies. Proposes two semantic relatedness measures to tackle the problem. A semantic relatedness measure based on traditional knowledge sources can be adapted. Wikipedia is a better knowledge source compared to traditional knowledge sources. A domain-specific semantic relatedness measure built on top of Wikipedia suits well for suggesting transfer course equivalencies. Provides a human judgment data set over 32 pairs of courses: http://bit.ly/semcourse.
  • 42. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesPublished Literature Using Semantic Distance to Automatically Suggest Transfer Course Equivalencies Beibei Yang and Jesse M. Heines ACL-HLT 2011: Proceedings of the Sixth Workshop on Innovative Use of NLP for Building Educational Applications (BEA-6) Association for Computational Linguistics Domain-Specific Semantic Relatedness from Wikipedia: Can a Course be Transferred? Beibei Yang and Jesse M. Heines NAACL-HLT 2012 Student Research Workshop
  • 43. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesReferencesBibliography I Alexander Budanitsky and Graeme Hirst. Evaluating Wordnet-based measures of lexical semantic relatedness. Computational Linguistics, 32:13–47, 2006. Curt Burgess, Kay Livesay, and Kevin Lund. Explorations in context space: words, sentences, discourse. Discourse Processes, 25:211–257, 1998. Danushka Bollegala, Yutaka Matsuo, and Mitsuru Ishizuka. Measuring semantic similarity between words using web search engines. In Proceedings of the 16th International Conference on World Wide Web, pages 757–766, New York, NY, USA, 2007. ACM. Rudi L. Cilibrasi and Paul M. B. Vitanyi. The google similarity distance. IEEE Transactions on Knowledge and Data Engineering, 19:370–383, 2007. Evgeniy Gabrilovich and Shaul Markovitch. Computing semantic relatedness using Wikipedia-based explicit semantic analysis. In Proceedings of the 20th International Joint Conference on AI, 2007. Evgeniy Gabrilovich and Shaul Markovitch. Wikipedia-based semantic interpretation for NLP. Journal of Artificial Intelligence Research, 34:443–498, 2009. Frederick Hayes-Roth, Donald A. Waterman, and Douglas B. Lenat. Building expert systems. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1983.
  • 44. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesReferencesBibliography II Graeme Hirst and David St-Onge. WordNet: An electronic lexical database, chapter Lexical chains as representations of context for the detection and correction of malapropisms, pages 305–332. The MIT Press, Cambridge, MA, 1998. Hideki Kozima and Teiji Furugori. Similarity between words computed by spreading activation on an english dictionary. In Proceedings of the 6th conference on European chapter of the Association for Computational Linguistics, EACL ’93, pages 232–239, Stroudsburg, PA, USA, 1993. Association for Computational Linguistics. Yuhua Li, Zuhair A. Bandar, and David McLean. An approach for measuring semantic similarity between words using multiple information sources. IEEE Transactions on Knowledge and Data Engineering, pages 871–882, 2003. Claudia Leacock and Martin Chodorow. Combining local context and WordNet similarity for word sense identification, pages 265–283. The MIT Press, Cambridge, MA, 1998. Thomas K Landauer, Peter W. Foltz, and Darrell Laham. An introduction to latent semantic analysis. Discourse Processes, 25(2-3):259–284, 1998. Yuhua Li, David McLean, Zuhair A. Bandar, James D. O’Shea, and Keeley Crockett. Sentence similarity based on semantic nets and corpus statistics. IEEE Transactions on Knowledge and Data Engineering, 18(8):1138–1150, 2006.
  • 45. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesReferencesBibliography III Jane Morris and Graeme Hirst. Lexical cohesion computed by thesaural relations as an indicator of the structure of text. Computational Linguistics, 17(1):21–48, March 1991. Distributional measures of concept-distance: A task-oriented evaluation, Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, 2006. Saif Mohammad. Measuring Semantic Distance Using Distributional Profiles of Concepts. PhD thesis, University of Toronto, Toronto, Canada, 2008. Simone Paolo Ponzetto and Michael Strube. Knowledge derived from Wikipedia for computing semantic relatedness. Journal of Artificial Intelligence Research, 30:181–212, October 2007. Philip Resnik. Using information content to evaluate semantic similarity in a taxonomy. In Proceedings of the 14th international joint conference on Artificial intelligence, volume 1 of IJCAI’95, pages 448–453, San Francisco, CA, USA, 1995. Morgan Kaufmann Publishers Inc. Gerard Salton and Christopher Buckley. Term weighting approaches in automatic text retrieval. Information Processing and Management, 24:513–523, August 1988. Mehran Sahami and Timothy D. Heilman. A web-based kernel function for measuring the similarity of short text snippets. In Proceedings of the 15th International Conference on the World Wide Web, pages 377–386, New York, NY, USA, 2006. ACM.
  • 46. Introduction Knowledge Sources Related Work First Approach Second Approach Summary ReferencesReferencesBibliography IV Peter D. Turney. Mining the web for synonyms: PMI-IR versus LSA on TOEFL. In Luc De Raedt and Peter A. Flach, editors, ECML, volume 2167 of Lecture Notes in Computer Science, pages 491–502. Springer, 2001. Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. In Proceedings 32nd Annual Meeting on Association for Computational Linguistics, pages 133–138, 1994. Dongqiang Yang and David M. W. Powers. Measuring semantic similarity in the taxonomy of wordnet. In Proceedings of the 28th Australasian Conference on Computer Science, volume 38, pages 315–322, Darlinghurst, Australia, 2005. Australian Computer Society, Inc.