TOEFL Synonym Questions (State of the art)
- TOEFL = Test of English as a Foreign Language
- 80 multiple-choice synonym questions; 4 choices per question
- the TOEFL questions are available on request by contacting LSA Support at CU Boulder, the people who manage the LSA web site at Colorado
- introduced in Landauer and Dumais (1997) as a way of evaluating algorithms for measuring degree of similarity between words
- subsequently used by many other researchers
Stem: levied Choices: (a) imposed (b) believed (c) requested (d) correlated Solution: (a) imposed
Table of results
|Algorithm||Reference for algorithm||Reference for experiment||Type||Correct||95% confidence|
|RES||Resnik (1995)||Jarmasz and Szpakowicz (2003)||Hybrid||20.31%||12.89–31.83%|
|LC||Leacock and Chodrow (1998)||Jarmasz and Szpakowicz (2003)||Lexicon-based||21.88%||13.91–33.21%|
|LIN||Lin (1998)||Jarmasz and Szpakowicz (2003)||Hybrid||24.06%||15.99–35.94%|
|Random||Random guessing||1 / 4 = 25.00%||Random||25.00%||15.99–35.94%|
|JC||Jiang and Conrath (1997)||Jarmasz and Szpakowicz (2003)||Hybrid||25.00%||15.99–35.94%|
|LSA||Landauer and Dumais (1997)||Landauer and Dumais (1997)||Corpus-based||64.38%||52.90–74.80%|
|Human||Average non-English US college applicant||Landauer and Dumais (1997)||Human||64.50%||53.01–74.88%|
|DS||Pado and Lapata (2007)||Pado and Lapata (2007)||Corpus-based||73.00%||62.72-82.96%|
|PMI-IR||Turney (2001)||Turney (2001)||Corpus-based||73.75%||62.72–82.96%|
|PairClass||Turney (2008)||Turney (2008)||Corpus-based||76.25%||65.42-85.06%|
|HSO||Hirst and St.-Onge (1998)||Jarmasz and Szpakowicz (2003)||Lexicon-based||77.91%||68.17–87.11%|
|JS||Jarmasz and Szpakowicz (2003)||Jarmasz and Szpakowicz (2003)||Lexicon-based||78.75%||68.17–87.11%|
|PMI-IR||Terra and Clarke (2003)||Terra and Clarke (2003)||Corpus-based||81.25%||70.97–89.11%|
|CWO||Ruiz-Casado et al. (2005)||Ruiz-Casado et al. (2005)||Web-based||82.55%||72.38–90.09%|
|PPMIC||Bullinaria and Levy (2007)||Bullinaria and Levy (2007)||Corpus-based||85.00%||75.26-92.00%|
|GLSA||Matveeva et al. (2005)||Matveeva et al. (2005)||Corpus-based||86.25%||76.73-92.93%|
|LSA||Rapp (2003)||Rapp (2003)||Corpus-based||92.50%||84.39-97.20%|
|ADW||Pilehvar et al. (2013)||Pilehvar et al. (2013)||WordNet graph-based (unsupervised)||96.25%||89.43-99.22%|
|PR||Turney et al. (2003)||Turney et al. (2003)||Hybrid||97.50%||91.26–99.70%|
|PCCP||Bullinaria and Levy (2012)||Bullinaria and Levy (2012)||Corpus-based||100.00%||96.32-100.00%|
Explanation of table
- Algorithm = name of algorithm
- Reference for algorithm = where to find out more about given algorithm
- Reference for experiment = where to find out more about evaluation of given algorithm with TOEFL questions
- Type = general type of algorithm: corpus-based, lexicon-based, hybrid
- Correct = percent of 80 questions that given algorithm answered correctly
- 95% confidence = confidence interval calculated using the Binomial Exact Test
- table rows sorted in order of increasing percent correct
- several WordNet-based similarity measures are implemented in Ted Pedersen's WordNet::Similarity package
- LSA = Latent Semantic Analysis
- PCCP = Principal Component vectors with Caron P
- PMI-IR = Pointwise Mutual Information - Information Retrieval
- PR = Product Rule
- PPMIC = Positive Pointwise Mutual Information with Cosine
- GLSA = Generalized Latent Semantic Analysis
- CWO = Context Window Overlapping
- DS = Dependency Space
- the performance of a corpus-based algorithm depends on the corpus, so the difference in performance between two corpus-based systems may be due to the different corpora, rather than the different algorithms
- the TOEFL questions include nouns, verbs, and adjectives, but some of the WordNet-based algorithms were only designed to work with nouns; this explains some of the lower scores
- some of the algorithms may have been tuned on the TOEFL questions; read the references for details
- Landauer and Dumais (1997) report scores that were corrected for guessing by subtracting a penalty of 1/3 for each incorrect answer; they report a score of 52.5% when this penalty is applied; when the penalty is removed, their performance is 64.4% correct
Bullinaria, J.A., and Levy, J.P. (2007). Extracting semantic representations from word co-occurrence statistics: A computational study. Behavior Research Methods, 39(3), 510-526.
Bullinaria, J.A., and Levy, J.P. (2012). Extracting semantic representations from word co-occurrence statistics: stop-lists, stemming, and SVD. Behavior Research Methods, 44(3):890-907.
Hirst, G., and St-Onge, D. (1998). Lexical chains as representation of context for the detection and correction of malapropisms. In C. Fellbaum (ed.), WordNet: An Electronic Lexical Database. Cambridge: MIT Press, 305-332.
Jarmasz, M., and Szpakowicz, S. (2003). Roget’s thesaurus and semantic similarity, Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP-03), Borovets, Bulgaria, September, pp. 212-219.
Jiang, J.J., and Conrath, D.W. (1997). Semantic similarity based on corpus statistics and lexical taxonomy. Proceedings of the International Conference on Research in Computational Linguistics, Taiwan.
Landauer, T.K., and Dumais, S.T. (1997). A solution to Plato's problem: The latent semantic analysis theory of the acquisition, induction, and representation of knowledge. Psychological Review, 104(2):211–240.
Leacock, C., and Chodorow, M. (1998). Combining local context and WordNet similarity for word sense identification. In C. Fellbaum (ed.), WordNet: An Electronic Lexical Database. Cambridge: MIT Press, pp. 265-283.
Lin, D. (1998). An information-theoretic definition of similarity. Proceedings of the 15th International Conference on Machine Learning (ICML-98), Madison, WI, pp. 296-304.
Matveeva, I., Levow, G., Farahat, A., and Royer, C. (2005). Generalized latent semantic analysis for term representation. Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP-05), Borovets, Bulgaria.
Pado, S., and Lapata, M. (2007). Dependency-based construction of semantic space models. Computational Linguistics, 33(2), 161-199.
Pilehvar, M.T., Jurgens D., and Navigli R. (2013). Align, disambiguate and walk: A unified approach for measuring semantic similarity. Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (ACL 2013), Sofia, Bulgaria.
Rapp, R. (2003). Word sense discovery based on sense descriptor dissimilarity. Proceedings of the Ninth Machine Translation Summit, pp. 315-322.
Resnik, P. (1995). Using information content to evaluate semantic similarity. Proceedings of the 14th International Joint Conference on Artificial Intelligence (IJCAI-95), Montreal, pp. 448-453.
Ruiz-Casado, M., Alfonseca, E. and Castells, P. (2005) Using context-window overlapping in Synonym Discovery and Ontology Extension. Proceedings of the International Conference Recent Advances in Natural Language Processing (RANLP-2005), Borovets, Bulgaria.
Terra, E., and Clarke, C.L.A. (2003). Frequency estimates for statistical word similarity measures. Proceedings of the Human Language Technology and North American Chapter of Association of Computational Linguistics Conference 2003 (HLT/NAACL 2003), pp. 244–251.
Turney, P.D. (2001). Mining the Web for synonyms: PMI-IR versus LSA on TOEFL. Proceedings of the Twelfth European Conference on Machine Learning (ECML-2001), Freiburg, Germany, pp. 491-502.
Turney, P.D., Littman, M.L., Bigham, J., and Shnayder, V. (2003). Combining independent modules to solve multiple-choice synonym and analogy problems. Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP-03), Borovets, Bulgaria, pp. 482-489.
Turney, P.D. (2008). A uniform approach to analogies, synonyms, antonyms, and associations. Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), Manchester, UK, pp. 905-912.