Assigning Stress to Out-Of-Vocabulary Words: Three Approaches

Transcript

1. AthenaRhythm Assigning Stress to Out-of-vocabulary Words Manex Agirrezabal1∗, Jeffrey Heinz2, Mans Hulden3, Bertol Arrieta1 (1) University of the Basque Country (UPV/EHU) (2) University of Delaware (3) University of Colorado Boulder July 23rd , 2014 The 2014 International Conference on Artificial Intelligence Las Vegas, Nevada, USA https://athenarhythm.googlecode.com https://zeuscansion.googlecode.com (*) At this moment, the first author is doing an internship in the University of Delaware 1/24, https://athenarhythm.googlecode.com
2. AthenaRhythm ... before we start ... 2/24, https://athenarhythm.googlecode.com
3. AthenaRhythm ... before we start ... Scansion Determine the rhythmic nature of lines of verse. Example and the mighty Mudjekeewis →´ and the | m´ ıghty | M´ udje|k´ eewis 2/24, https://athenarhythm.googlecode.com
4. AthenaRhythm ... before we start ... Scansion Determine the rhythmic nature of lines of verse. Example and the mighty Mudjekeewis →´ and the | m´ ıghty | M´ udje|k´ eewis Trochee: − 2/24, https://athenarhythm.googlecode.com
5. AthenaRhythm ... before we start ... Scansion Determine the rhythmic nature of lines of verse. Example and the mighty Mudjekeewis →´ and the | m´ ıghty | M´ udje|k´ eewis Trochee: − Four trochees Trochaic tetrameter 2/24, https://athenarhythm.googlecode.com
6. AthenaRhythm ... what you’ll see in this presentation ... 3/24, https://athenarhythm.googlecode.com
7. AthenaRhythm ... what you’ll see in this presentation ... A prerequsite problem: how to determine the location of stress in words that are not in the dictionary (out-of-vocabulary words) 3/24, https://athenarhythm.googlecode.com
8. AthenaRhythm ... what you’ll see in this presentation ... A prerequsite problem: how to determine the location of stress in words that are not in the dictionary (out-of-vocabulary words) Three approaches to predict the primary stress location: 3/24, https://athenarhythm.googlecode.com
9. AthenaRhythm ... what you’ll see in this presentation ... A prerequsite problem: how to determine the location of stress in words that are not in the dictionary (out-of-vocabulary words) Three approaches to predict the primary stress location: Word Similarity Linguistic Generalizations Machine Learning 3/24, https://athenarhythm.googlecode.com
10. AthenaRhythm ... what you’ll see in this presentation ... A prerequsite problem: how to determine the location of stress in words that are not in the dictionary (out-of-vocabulary words) Three approaches to predict the primary stress location: Word Similarity Linguistic Generalizations Machine Learning The best results with Linguistic Generalizations, but Machine Learning is not so far. 3/24, https://athenarhythm.googlecode.com
11. AthenaRhythm Outline 1 Background 2 Technical details: Corpus & Software 3 Task 4 Approaches Similarity approach Linguistic approach Machine Learning approach 5 Results 6 Discussion & Future Work 4/24, https://athenarhythm.googlecode.com
12. AthenaRhythm Background 1 Background 2 Technical details: Corpus & Software 3 Task 4 Approaches Similarity approach Linguistic approach Machine Learning approach 5 Results 6 Discussion & Future Work 5/24, https://athenarhythm.googlecode.com
13. AthenaRhythm Background FACTS: The location of primary stress in English words is a key element of successful scansion of English poetry. Scansion systems fail when they can’t get the lexical stress of the words ZeuScansion: A tool for scansion of English poetry (Agirrezabal et al., 2013) Similarity approach for unknown words We have developed other two approaches 6/24, https://athenarhythm.googlecode.com
14. AthenaRhythm Technical details: Corpus & Software 1 Background 2 Technical details: Corpus & Software 3 Task 4 Approaches Similarity approach Linguistic approach Machine Learning approach 5 Results 6 Discussion & Future Work 7/24, https://athenarhythm.googlecode.com
15. AthenaRhythm Technical details: Corpus & Software Corpus NETtalk pronunciation dictionary Software Foma: Finite-state compiler and C library Phonetisaurus: Grapheme-to-Phoneme converter LIBLINEAR: Library for Large Linear Classification LIBSVM: Library for Support Vector Machines 8/24, https://athenarhythm.googlecode.com
16. AthenaRhythm Task 1 Background 2 Technical details: Corpus & Software 3 Task 4 Approaches Similarity approach Linguistic approach Machine Learning approach 5 Results 6 Discussion & Future Work 9/24, https://athenarhythm.googlecode.com
17. AthenaRhythm Task Goal Given an out-of-vocabulary word, infer the location of the primary stress 10/24, https://athenarhythm.googlecode.com
18. AthenaRhythm Task Goal Given an out-of-vocabulary word, infer the location of the primary stress Mudjekeewis Where is the primary stress? 10/24, https://athenarhythm.googlecode.com
19. AthenaRhythm Task Goal Given an out-of-vocabulary word, infer the location of the primary stress M´ udjekeewis Where is the primary stress? 10/24, https://athenarhythm.googlecode.com
20. AthenaRhythm Task Goal Given an out-of-vocabulary word, infer the location of the primary stress Mudj´ ekeewis Where is the primary stress? 10/24, https://athenarhythm.googlecode.com
21. AthenaRhythm Task Goal Given an out-of-vocabulary word, infer the location of the primary stress Mudjek´ e´ ewis Where is the primary stress? 10/24, https://athenarhythm.googlecode.com
22. AthenaRhythm Task Goal Given an out-of-vocabulary word, infer the location of the primary stress Mudjekeew´ ıs Where is the primary stress? 10/24, https://athenarhythm.googlecode.com
23. AthenaRhythm Task Goal Given an out-of-vocabulary word, infer the location of the primary stress Mudjekeewis Where is the primary stress? The system has to learn how to locate the stress 10/24, https://athenarhythm.googlecode.com
24. AthenaRhythm Approaches 1 Background 2 Technical details: Corpus & Software 3 Task 4 Approaches Similarity approach Linguistic approach Machine Learning approach 5 Results 6 Discussion & Future Work 11/24, https://athenarhythm.googlecode.com
25. AthenaRhythm Approaches Similarity approach Similarity approach Our thought Similarly spelled words have the same lexical stress. Example I prophesy there’ll be a row 12/24, https://athenarhythm.googlecode.com
26. AthenaRhythm Approaches Similarity approach Similarity approach Our thought Similarly spelled words have the same lexical stress. Example prophesy 12/24, https://athenarhythm.googlecode.com
27. AthenaRhythm Approaches Similarity approach Similarity approach Our thought Similarly spelled words have the same lexical stress. Example prop|hesy 12/24, https://athenarhythm.googlecode.com
28. AthenaRhythm Approaches Similarity approach Similarity approach Our thought Similarly spelled words have the same lexical stress. Example prop|hecy 12/24, https://athenarhythm.googlecode.com
29. AthenaRhythm Approaches Similarity approach Similarity approach Our thought Similarly spelled words have the same lexical stress. Example prophecy We can get it from the dictionary 12/24, https://athenarhythm.googlecode.com
30. AthenaRhythm Approaches Similarity approach Similarity approach Our thought Similarly spelled words have the same lexical stress. Example prophecy − − We can get it from the dictionary 12/24, https://athenarhythm.googlecode.com
31. AthenaRhythm Approaches Linguistic approach Linguistic approach Linguistic chain: Grapheme-to-phoneme (G2P) conversion (Novak et al., 2012) Syllabification procedure (Hulden, 2006) Stress-assignment rules (Hayes, 1995) Generalization: Heavy syllables tend to attract stress 13/24, https://athenarhythm.googlecode.com
32. AthenaRhythm Approaches Linguistic approach Linguistic approach Stress assignment rules dialectal dYxlEktL dY.x.lEk.tL dY.x.'lEk.tL 14/24, https://athenarhythm.googlecode.com
33. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference 15/24, https://athenarhythm.googlecode.com
34. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference Feature extraction {#r}:1, {ef}:1, {fe}:1, {er}:1 {en}:1, {nc}:1, {ce}:1, {e#}:1 {re}:2 15/24, https://athenarhythm.googlecode.com
35. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference Feature extraction {#r}:1, {ef}:1, {fe}:1, {er}:1 {en}:1, {nc}:1, {ce}:1, {e#}:1 {re}:2 , class: < − > 15/24, https://athenarhythm.googlecode.com
36. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference Feature extraction {#r}:1, {ef}:1, {fe}:1, {er}:1 {en}:1, {nc}:1, {ce}:1, {e#}:1 {re}:2 , class: < − > Learning linear SVMs, non-linear SVMs, different kernels 15/24, https://athenarhythm.googlecode.com
37. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference Feature extraction {#r}:1, {ef}:1, {fe}:1, {er}:1 {en}:1, {nc}:1, {ce}:1, {e#}:1 {re}:2 , class: < − > Learning linear SVMs, non-linear SVMs, different kernels Hyperparameter optimization Grid-search for C and γ params. 15/24, https://athenarhythm.googlecode.com
38. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference Feature extraction {#r}:1, {ef}:1, {fe}:1, {er}:1 {en}:1, {nc}:1, {ce}:1, {e#}:1 {re}:2 , class: < − > Learning linear SVMs, non-linear SVMs, different kernels Hyperparameter optimization Grid-search for C and γ params. The best 15/24, https://athenarhythm.googlecode.com
39. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference Feature extraction {#r}:1, {ef}:1, {fe}:1, {er}:1 {en}:1, {nc}:1, {ce}:1, {e#}:1 {re}:2 , class: < − > Learning linear SVMs, non-linear SVMs, different kernels Hyperparameter optimization Grid-search for C and γ params. The best C-Support Vector Classifier 15/24, https://athenarhythm.googlecode.com
40. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference Feature extraction {#r}:1, {ef}:1, {fe}:1, {er}:1 {en}:1, {nc}:1, {ce}:1, {e#}:1 {re}:2 , class: < − > Learning linear SVMs, non-linear SVMs, different kernels Hyperparameter optimization Grid-search for C and γ params. The best C-Support Vector Classifier with RBF kernel 15/24, https://athenarhythm.googlecode.com
41. AthenaRhythm Approaches Machine Learning approach Machine Learning approach Example reference Feature extraction {#r}:1, {ef}:1, {fe}:1, {er}:1 {en}:1, {nc}:1, {ce}:1, {e#}:1 {re}:2 , class: < − > Learning linear SVMs, non-linear SVMs, different kernels Hyperparameter optimization Grid-search for C and γ params. The best C-Support Vector Classifier with RBF kernel and C = 1024 and γ = 0.0078125 15/24, https://athenarhythm.googlecode.com
42. AthenaRhythm Results 1 Background 2 Technical details: Corpus & Software 3 Task 4 Approaches Similarity approach Linguistic approach Machine Learning approach 5 Results 6 Discussion & Future Work 16/24, https://athenarhythm.googlecode.com
43. AthenaRhythm Results Evaluation method: Cross-validation for Similarity approach and Machine Learning approach Whole corpus for the Linguistic approach (Expert system) 17/24, https://athenarhythm.googlecode.com
44. AthenaRhythm Results Approach Accuracy SIM 0.5843 SIM-OO 0.6777 LING 0.7362 SVM 0.7098 SIM: Similarity approach SIM-OO: Similarity approach with optimal ordering of transducers LING: Linguistic approach SVM: Support Vector Machines (with the BEST parameters, C = 1024 and γ = 0.0078125) 18/24, https://athenarhythm.googlecode.com
45. AthenaRhythm Results Approach Accuracy SIM 0.5843 SIM-OO 0.6777 LING 0.7362 SVM 0.7098 SIM: Similarity approach SIM-OO: Similarity approach with optimal ordering of transducers LING: Linguistic approach SVM: Support Vector Machines (with the BEST parameters, C = 1024 and γ = 0.0078125) 18/24, https://athenarhythm.googlecode.com
46. AthenaRhythm Discussion & Future Work 1 Background 2 Technical details: Corpus & Software 3 Task 4 Approaches Similarity approach Linguistic approach Machine Learning approach 5 Results 6 Discussion & Future Work 19/24, https://athenarhythm.googlecode.com
47. AthenaRhythm Discussion & Future Work We have presented three approaches for predicting primary stress in out-of-vocabulary words. 20/24, https://athenarhythm.googlecode.com
48. AthenaRhythm Discussion & Future Work We have presented three approaches for predicting primary stress in out-of-vocabulary words. Important part of ZeuScansion 20/24, https://athenarhythm.googlecode.com
49. AthenaRhythm Discussion & Future Work We have presented three approaches for predicting primary stress in out-of-vocabulary words. Important part of ZeuScansion (The automatic scansion system). 20/24, https://athenarhythm.googlecode.com
50. AthenaRhythm Discussion & Future Work We have presented three approaches for predicting primary stress in out-of-vocabulary words. Important part of ZeuScansion (The automatic scansion system). Best results: Linguistic rules and Machine Learning 20/24, https://athenarhythm.googlecode.com
51. AthenaRhythm Discussion & Future Work We have presented three approaches for predicting primary stress in out-of-vocabulary words. Important part of ZeuScansion (The automatic scansion system). Best results: Linguistic rules and Machine Learning These implementations are released under GNU GPL license at athenarhythm.googlecode.com 20/24, https://athenarhythm.googlecode.com
52. AthenaRhythm Discussion & Future Work Future Work: Include part-of-speech information More feature sets (Machine Learning approach) Improve the linguistic rules, expecially disyllables Apply these approaches in other languages 21/24, https://athenarhythm.googlecode.com
53. AthenaRhythm Acknowledgments Thanks to: 22/24, https://athenarhythm.googlecode.com
54. AthenaRhythm Acknowledgments Thanks to: Department of Linguistics & Cognitive Sciences (University of Delaware) 22/24, https://athenarhythm.googlecode.com
55. AthenaRhythm Acknowledgments Thanks to: Department of Linguistics & Cognitive Sciences (University of Delaware) University of the Basque Country (UPV/EHU) 22/24, https://athenarhythm.googlecode.com
56. AthenaRhythm Acknowledgments Thanks to: Department of Linguistics & Cognitive Sciences (University of Delaware) University of the Basque Country (UPV/EHU) NSF grant #1123692 to the second author. 22/24, https://athenarhythm.googlecode.com
57. AthenaRhythm Questions? 23/24, https://athenarhythm.googlecode.com
58. AthenaRhythm Assigning Stress to Out-of-vocabulary Words Manex Agirrezabal1∗, Jeffrey Heinz2, Mans Hulden3, Bertol Arrieta1 (1) University of the Basque Country (UPV/EHU) (2) University of Delaware (3) University of Colorado Boulder July 23rd , 2014 The 2014 International Conference on Artificial Intelligence Las Vegas, Nevada, USA https://athenarhythm.googlecode.com https://zeuscansion.googlecode.com (*) At this moment, the first author is doing an internship in the University of Delaware 24/24, https://athenarhythm.googlecode.com
59. AthenaRhythm Hayes, B. (1995). Metrical stress theory: Principles and case studies. University of Chicago Press. Hulden, M. (2006). Finite-state syllabification. Finite-State Methods and Natural Language Processing, pages 86–96. Novak, J., Minematsu, N., and Hirose, K. (2012). Wfst-based Grapheme-to-Phoneme conversion: Open source tools for alignment, model-building and decoding. Finite-State Methods and Natural Language Processing. 24/24, https://athenarhythm.googlecode.com