ItGraSyll: A Computational Analysis of Graphical Syllabification and Stress Assignment in Italian

Abstract

In this paper we build a dataset of Italian graphical syllables (called ItGraSyll). We perform quantitative and qualitative analyses on the syllabification and stress assignment in Italian. We propose a machine learning model, based on deep-learning techniques, for automatically inferring syllabification and stress assignment. For stress prediction we report 94.45% word-level accuracy, and for syllabification we report 98.41% word-level accuracy and 99.82% hyphen-level accuracy. “prosodic revolution” [10] from Latin to the Romance languages - including syncope (the loss of an intermediate Word syllabification and syllable analysis are two related syllable) and apocope (the loss of the final syllable) at a issues of great importance in the study of language (writ- large scale - has led to major changes, but their weight is ten or spoken). These topics have attracted a large cat- diferent from one idiom to another: while the Western egory of researchers, from pure linguists, in phonetics, Romance languages manifest highly evident diferences to psycholinguists, computer scientists, speech thera- from the Latin phonological and prosodic system, and the pists, etc. Thus, the syllable plays an important role in Eastern languages are considered to be most conservative language learning and acquisition, speech recognition, from this point of view, Italian seems to be in between speech production [1, 2], language similarity [3], in text [10]. On the other hand, in Latin, the relation between comprehensibility (Kincaid-Flesch formula [4]), in speech stress and quantity grew stronger, thus short stressed therapy, in poetry analysis [5, 6], etc. Each language has vowels progressively gained length. It is noteworthy that its own way of grouping sounds into syllables and its own this situation is best preserved in Italian, and not in the rules for dividing words into syllables. Linguistically, the Eastern Romance idioms: thus, in Italian stress cannot syllable represents "the smallest phonetic trance likely skip a heavy penultimate syllable, and stress cannot fall to receive an accent and only one" [7], and the syllabic further back than the antepenultimate syllable, a twofold cut is seen by De Saussure [8] on the border between the characteristic feature of the Latin prosodic system. This implosion and the explosion of the spoken sound: "If in is why we are taking Italian as a starting point for a largera chain of sounds one goes from implosion to explosion, scale study, oriented towards all Romance languages. The one obtains a particular efect which is the indication of main diference between Latin and its modern descenthe boundary of the syllable". dants is that Latin stress was quantity- sensitive, leading The analysis of the words' syllabic structure also plays thus to the following rule: in polysyllabic words, stress an important part in historical linguistics [9], not only fell on a heavy penultimate (meaning, containing a long in diachronic phonetics and phonology, but also in lexi- vowel), otherwise on the antepenultimate. Due to the cology. Romance comparative linguistics, in particular, collapse of vowel quantity as a distinctive feature in the still needs a detailed overview of this aspect, as syllable, vocalic system, no Romance language has retained the segmentation and prosody can give strong account on Latin stress rule as such [10]. As, from a statistic point of phonetic changes that haven't been explained yet. The view, the greatest part of the Romance lexicon is represented by penultimate stressed words, a basic automatic mechanism would assign penultimate stress by default, whereas for both final and antepenultimate stress, the machine (as well as, not in a few cases, non-native speakers) would need further specification. As a consequence of the loss of Latin vowel quantity, Romance stress has ceased to be completely predictable. That is, partially,

Keywords

• eol>syllabification
• stress assignment
• Italian

1. Introduction

CLiC-it 2024: Tenth Italian Conference on Computational Linguistics, Dec 04 — 06, 2024, Pisa, Italy * Corresponding author. $ [e-mail] (L. P. Dinu); [e-mail] (B. Iordache); [e-mail] (B. Guita); [e-mail] (S. Georgescu); [e-mail] (A. Cristea) © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License why in the majority of the traditional Romance comparaAttribution 4.0 International (CC BY 4.0). tive or historical grammars, there is no specific section other linguistic factors that those rules take into account. devoted to syllabification [ 11], or, if there is, it focuses For example, a rule that is present in many languages either on general prosodic features [12], or on the vowel distinguishes between a vowel and a semivowel, but the evolution depending on its presence in an open or closed computer is not able to easily recognize when the same syllable [13]. The lack of a section dedicated to syllab- sign has the value of a vowel and when it is a semivowel. ification is also common in the historical grammars of Because of this, rule-based adaptations of syllabification Italian [14, 11, 15]. We will focus in this research only systems [26] generally have higher errors, and many lanon written form of words, so we will investigate only guages do not have an automatic syllabification system the graphical syllabification and stress. By focusing on yet (for example, in the Python library, only a few lanthe graphical syllabification and stress in Italian, we aim guages have syllabification). The last few decades have to take a step forward towards the complete evaluation brought the first data-driven syllabification systems. of the prosodic changes that took place in the transition However, in order to build such a system, training from Latin to the Romance languages, and their influence data is needed, and there are many cases in which the on the Romance phonetics and phonology. A machine- available data do not cover the whole language, and thus learning model, capable of automatically inferring graph- the systems have diferent results when the test corpus ical syllabification and stress assignment, along with the is changed. purpose of creating a data-base containing the quanti- Starting with these remarks, our main contributions tative and qualitative description of syllabification and are: stress in the Romance languages, could be the first important task in the greater challenge of tracing the simi- • We propose ItGraSyll (Italian graphical syllables), larities and diferences between the Romance languages a dataset of 114, 503 Italian words, in orthoand, more important, between Romance and Latin. From graphic form, containing annotations for their ora typological point of view, the study of syllabification thographic syllabification and stress placement 1 and stress can shed a new light on the universal features • We perform quantitative and qualitative analyses that, by defining our phonoarticulatory and phonoacous- of the previously built dataset. tic apparatus, have guided the languages’ development • We analyze stress placement in the context of the and change. Given the promising results of this analysis, Italian syllables. the present study can establish the basis of a research of • We propose an automatic system of syllabification the syllable in other languages, either linguistically or for Italian words. typologically related to Italian.

One of the studies that address automatic syllabification in Italian belongs to Bigi and Petrone [16], who 2. Quantitative Analysis proposed a tool that performs rule-based automatic segmentation. Adsett and Marchand [17] and Adsett et al. In this section we perform various measurements regard[18] investigated whether data-driven approaches out- ing the syllables and stress placement of Italian written perform rule-based approaches for a language with a words and analyze the results. We perform, on Italian, low syllabic complexity, such as Italian. The authors an investigation similar to a previous investigations conreached the conclusion that even in this case data-driven ducted on Romanian by Dinu and Dinu [27], Dinu and systems are the more appropriate approach. In terms of Dinu [ 28 ]. machine learning, the tasks of automatically inferring syllable boundaries and predicting stress assignment can be 2.1. Data naturally framed as sequence labeling problems. While automatic syllabification has received more attention re- We build a dataset of Italian words starting from the cently [19, 20, 21, 22, 23, 24], stress placement has not online version of Dizionario italiano De Mauro,2 which been investigated as much [25]. provides information regarding graphical syllabification

Given the complexity of syllable applications and word and stress placement for the Italian vocabulary. Stressed syllabification, the presence of electronic resources dedi- syllables are also shown by having accents on the domicated to them becomes a necessity. While native speakers nant vowel. Going further, this dataset will be referred of a language generally do not have great dificulty in to as ItGraSyll. spelling words, the same cannot be said of those who We performed several pre-processing steps. We learn a foreign language who often tend to apply their cleaned the resulted dataset by removing duplicates, preown rules to foreign words, and problems arise in au- fixes and sufixes in order to remain with the base word; tomatic syllabification. This is because the rules of syllabification are linguistic rules, and they cannot always 1Ththtepsd:/a/tnalspe.utnisibauvc.ariola/rbelseofuorrcerse.shetamrlc#hitgpruarspyollses upon request at: be easily modeled by the computer when there are no 2https://dizionario.internazionale.it/ abbreviations and unwanted punctuation marks such as dots, commas, apostrophes and dashes were also excluded so we can correctly process each word and its syllable division. Finally, the dataset consists of 114, 503 words in orthographic form having between one and eleven syllables. The distribution of words per number of syllables is represented in Table 1.

#syll.

#words

Examples

2.2. Syllables

2.3. Syllable Structure

We identified a total of 67 diferent consonant-vowel structures. The most frequent 7 structures cover almost 97% of the total. Depending on the type-token ratio,

the most frequent consonant-vowel structures are the following: a) for the type syllables: cvc (25%), ccvc (20.9%), cvvc (7.79%). b) for the token syllables: cv (58%), cvc (15%), ccv (7%), cvv (4.74%) and v (4.32%). Moreover, we observe that the cv structure corresponds to 40 out of the most frequent 50 syllables from the dataset.

2.4. Stress Placement

We identified a total of 2,883 stressed syllables (type syl

lables). So, 847 syllables are never stressed. The most frequent 20 stressed syllables are represented in Table 3. We observe that the most frequent stressed syllable (men) has a very high stress ratio (90%) when we compare the stressed occurrences with all its occurrences (stressed and unstressed) in our database. While in the top 20 of all syllables, men is the only syllable of length 3 (on the 14th position), for stressed syllables there are a couple of other syllables with a length greater than 2 (zio on position 6 with 34% stress ratio, gia on position 19 with 65% stress ratio).

We investigate stress placement with regard to syllable structure and we provide in Table 4 the percentages of words having the stress placed on diferent positions (for top 5), counting syllables from the beginning and from the end of the words as well. We observe that in most cases the stress is placed on the second to last syllable.

Stress ratio (%) 2.5. Syllables’ Usage 100 cover 74% and the most frequent 150 syllables (i.e. 4% of # ) cover 80% of # . Over this number, the percentage of coverage rises slowly. 2,281 (61%) syllables of type syllables occur less then 10 times, and 1,174 syllables occur only once (hapax legomena). 2.5.2. Stressed Syllables

A similar trend can be observed also for the stressed syl

lables. Further, we notice that the most frequent syllables cover a wide ratio of the total syllable frequency. For example, the 10 most frequent stressed syllable represent 31% of the total of stressed syllables, the top 50 syllables, 60% and the top 200 syllables, 81% of the token syllables. The values are plotted in Figure 1, for all syllables and for stressed syllables.

0.8 0.7 e0.6 g a rve oC0.5 0.4 0.3

Type al syl ables stressed syl ables 25 50 75Number of syl a1b2le5s 150 175 200

100

This results proves that the law is true for Italian too,

a very small number of syllables cover a large part from Italian language (there are necessary only 150 syllables to cover 80% from language).

3. Minimum Efort Laws

In this section we discuss two minimum efort laws that have been previously investigated for other languages and verify whether they apply for Italian as well.

2.5.1. General Syllables The syllables have a less intuitive behaviour, usually a small number of syllables cover a large part from a lan- 3.1. Chebanow guage. This is valuable for a large category of natural Denoting by () the frequency of a word having n languages, including English, Dutch, Romanian [ 28 ], Ko- syllables and by = ∑︀ ()/ ∑︀ () the average rean, Chinese, etc. We investigate here if this empirical length (measured in syllables) of the words, Chebanow law is also applicable to Italian. We made this investiga- [ 29 ] proposed the following law between the average tion both on stressed and general syllables. and the probability of occurrences () of the words having n syllables:

The most frequent 30 Italian syllables (when stress place

ment is disregarded) cover almost 50% of # , the most frequent 50 syllables cover 61%, the most frequent For Italian, = 4.226.

() = ( − 1)− 1 ( − 1)! * 1− (1) (a) The probability distribution of the

length of words.

In Figures 2a and 2b we plot the probability distribution of the length of words (in syllables) – the practical and theoretical representations.

We observe that the two curves have comparable We further investigate how a deep-learning model can aushapes, with a more prominent peak for the probabil- tomatically infer the syllabification and stress assignment ity distribution in Figure 2a; this peak can be influenced of Italian words, given their orthographic representation. by the fact that it is determined based on all the words in the dictionary, where many 4-syllable words are present.

4. Automatic Syllabification and

Stress Assignment

4.1. Methodology

3.2. Menzerath Both tasks can be defined in terms of a sequence labelling problem, strategy which was previously successMenzerath’s law – later generalized by the Menzerath- ful used for Romanian[ 31, 32 ]. Let us consider, for exAltmann law [ 30 ] – states that the bigger the number of ample, the word medaglione (the Italian translation of syllables in a word, the lesser the number of phonemes the word "locket"). For syllabification we can label each composing these syllables. In other words, Menzerath’s letter from the word either with the label 1, denoting law expresses a negative correlation between the length that a syllable starts from that letter, or with the label of a word in syllables and the lengths in phonemes of its 0, meaning the respective letter is not the first letter in constitutive syllables. In cognitive economy terms, this its syllable. Similarly, for identifying the stressed vowel, means that the more complex a linguistic construct, the we can label its position with a 1 and all other letters smaller its constituents. The law is expressed as follows: are assigned the label 0. We thus obtain for our exam = − (2) speleq utheencseeq0u0e0n0c0e0100110001f0o0r0s1t0refsosrpsryelldaibcitficioatnio(ni.ea.nmd et-hdeawhere is the syllable length (the size of the constituent), gliò-ne, the o vowel is stressed). is the number of syllables per word (the size of the lin- With these definitions, we can now construct machine guistic construct), and , , are empirical parameters. learning models for labelling the character sequences. Figure 2c shows that the law is satisfied for Italian. The model we propose is a recurrent neural network based on Gated Recurrent Units (GRU) [ 33 ]. The model architecture is comprised from the following components: • a character embedding layer, producing 64- 4.2. Results Anaysis dimensional vectors for each unique character

Table 5 contains the metrics computed on the test set, • a stacked bidirectional GRU, with 3 layers and a using the models trained for syllabification (both with 128-dimensional hidden state; a 0.2-rate dropout and without stress markers) and the model trained for applied after each of the first two layers predicting the stressed vowel. We obtained a remarkable • 0.5-rate dropout, after the last GRU layer, along hyphen accuracy of 99.74% for syllabification without with one-dimensional batch normalization the stress markers, and, when we add the stress markers, • a time-distributed fully-connected layer with 256 we obtained an increasing accuracy, obtaining 99.82%.

output nodes and ReLU activation Including the stress markers into the data used for syl• a linear layer that projects the 256-dimensional labification improved the metrics across the board, most vector into a single number, on which sigmoid notably with a ∼ 1% increase in word-level accuracy, activation is applied to infer the binary labels. which considering the large amount of data, and the high accuracy scores is a significant improvement ( 460 fewer

For training the models for both tasks, the dataset of syllabification mistakes as opposed to the approach that words is split into 50% training examples and 50% test excludes stress markers). Regarding the stress prediction, examples, unseen during training. we obtained an accuracy of 94.45%. Table 6 showcases a

The loss function computed for the prediction made series of wrong predictions generated by the models on for a word, regardless of the task on which the model the tests sets for stress assignment and syllabification. is trained, is the average of two terms: the first one is We also look into the accuracy scores computed for the average character-wise binary cross-entropy, while the test set, when it is bucketed based on the real number the second one is the root mean squared error computed of syllables of the test words. These results are shown between the vector of predicted labels and the ground- in Figure 3 and Table 7. For stress assignment, accutruth vector. The model is optimized using the Adam racy decreases to a global minimum for disyllabic words, optimizer [ 34 ], with a learning rate of 0.0003, no weight then starts to increase again with the number of syllables. decay, bath size of 32, and a LR scheduler that halves it For the syllabification task, including the stress markers every 5 epochs. The models are trained for 10-15 epochs. seems to outperform excluding them in most scenarios,

For the task of automatic syllabification, we wanted while both accuracies achieve a peak around the 5 syllato check if the presence of the stress markers afects the bles mark. This result seems to align with the distribution performance of the model. Because of that, we trained of syllables in the dataset, i.e. obtaining higher scores two models: the first one was trained using the spelling for the number of syllables with more examples. For of the words with the stress markers removed, while the stress assignment errors, we also investigate the placesecond one was trained with them included. ment of the predicted stressed syllable in relation with Stress Assignment Errors tthhee tsrtureesosnede (ssyelelaTbaleblaet8m). o9s5t.6o%neopfotshietioernrotorstmheislpelfatc,eodr to the right, while almost two thirds of the erroneous predictions placed the stress on the first syllable to the right of the correct one.

Predicted balanò femòre dolmèn tutòlo pùdico còrsia True bàlano fèmore dòlmen tùtolo pudìco corsìa Syllabification Errors

True mu-o-ne bion-da cli-en-te co-di-a-to ma-nu-brio spa-tria-to

Predicted muo-ne bi-on-da clien-te co-dia-to ma-nu-bri-o spa-tri-a-to

Syllabification (w/o SM)

5. Conclusions

In this paper we have investigated graphical syllabifica

tion and graphical stress assignment for Italian words. We have started by building ItGraSyll, a dataset of Italian graphical syllabified words, with stress annotations as well, on which we have performed several quantitative and qualitative analyses, including the verification of two minimum efort laws for the case of Italian. Finally, we have proposed a recurrent neural network machine learning model for automatic syllabification and stress assignment for Italian written words. For stress prediction we have obtained 94.45% word-level accuracy, and for syllabification we have obtained 98.41% word-level accuracy and 99.82% hyphen-level accuracy. In future work we intend to extend the analysis from dictionary level to corpus level and to investigate other languages as well.

Acknowledgments

We want to thank the reviewers for their useful suggestions. Research supported by the Ministry of Research,

Innovation and Digitization, CNCS/CCCDI UEFISCDI, SiRoLa project, number PN-IV-P1-PCE-2023-1701, Romania.

in Computer Science, Springer, 2005, pp. 785–788. 463. URL: https://doi.org/10.1016/j.csl.2009.02.004. URL: https://doi.org/10.1007/978-3-540-30586-6_88. doi:10.1016/j.csl.2009.02.004. doi:10.1007/978-3-540-30586-6\_88. [19] K. A. Rogova, K. Demuynck, D. V. Compernolle, Au[4] J. P. Kincaid, L. R. P. F. Jr., R. L. Rogers, B. S. Chissom, tomatic syllabification using segmental conditional Derivation of new readability formulas (Automated random fields, in: Computational Linguistics in the Readability Index, Fog Count and Flesch Reading Netherlands Journal, volume 3, 2013, pp. 34–48. Ease formula) for Navy enlisted personnel, Re- [20] L. P. Dinu, V. Niculae, O. Sulea, Romanian syllabsearch Branch Report, Millington, TN: Chief of ication using machine learning, in: I. Habernal, Naval Training, 1975. V. Matousek (Eds.), Text, Speech, and Dialogue [5] G. Marco, J. de la Rosa, J. Gonzalo, S. Ros, 16th International Conference, TSD 2013, Pilsen, E. González-Blanco, Automated Metric Analysis of Czech Republic, September 1-5, 2013. Proceedings, Spanish Poetry: Two Complementary Approaches, volume 8082 of Lecture Notes in Computer Science, IEEE Access 9 (2021) 51734–51746. Springer, 2013, pp. 450–456. [6] A. M. Ciobanu, L. P. Dinu, On the romanian [21] J. Krantz, M. W. Dulin, P. D. Palma, Languagerhyme detection, in: Proceedings of COLING 2012: Agnostic Syllabification with Neural Sequence LaDemonstration Papers, 2012, pp. 87–94. beling, 2019 18th IEEE International Conference [7] L. Hjelmslev, The syllable as a structural unit, in: On Machine Learning And Applications (ICMLA) the Proceedings of the 3rd International Congress (2019) 804–810. of Phonetic Sciences (Ghent), 1938, volume 266, [22] V. N. Vitale, L. Schettino, F. Cutugno, On incre1938. menting interpretability of machine learning mod[8] F. De Saussure, Course in general linguistics, els from the foundations: A study on syllabic speech

Columbia University Press, 2011. units, in: F. Boschetti, G. E. Lebani, B. Magnini, [9] D. Russo, The Notion of Syllable across History, N. Novielli (Eds.), Proceedings of the 9th Italian Theories and Analysis, Cambridge Scholars Pub- Conference on Computational Linguistics, Venice, lishing, 2016. Italy, November 30 - December 2, 2023, volume 3596 [10] M. Loporcaro, Syllable, segment and prosody, in: of CEUR Workshop Proceedings, CEUR-WS.org, 2023.

The Cambridge history of the Romance languages, URL: https://ceur-ws.org/Vol-3596/paper51.pdf . 2011, pp. 50–108. [23] O. Sulea, L. P. Dinu, B. Dumitru, Full inflec[11] W. Meyer-Lübke, Grammaire des langues romanes, tion learning using deep neural networks, in: volume 4, H. Welter, 1906. A. F. Gelbukh (Ed.), Computational Linguistics [12] M.-D. Glessgen, Linguistique romane: domaines and Intelligent Text Processing - 19th Internaet méthodes en linguistique française et romane, tional Conference, CICLing 2018, Hanoi, Vietnam, Armand Colin, 2007. March 18-24, 2018, Revised Selected Papers, Part [13] F. S. Miret, Fonética histórica, in: Manual de lingüís- I, volume 13396 of Lecture Notes in Computer Scitica románica, Ariel España, 2007, pp. 227–250. ence, Springer, 2018, pp. 408–415. URL: https://doi. [14] F. d’Ovidio, W. Meyer-Lübke, Grammatica storica org/10.1007/978-3-031-23793-5_33. doi:10.1007/ della lingua e dei dialetti italiani, volume 368, U. 978-3-031-23793-5\_33.

Hoepli, 1906. [24] M. Petrillo, F. Cutugno, A syllable segmentation al[15] G. Rohlfs, T. Franceschi, Grammatica storica della gorithm for english and italian., in: INTERSPEECH lingua italiana e dei suoi dialetti: Morfologia, (No 2003, 2003, pp. 2913–2916.

Title) (1968). [25] Q. Dou, S. Bergsma, S. Jiampojamarn, G. Kondrak, A [16] B. Bigi, C. Petrone, A generic tool for the automatic Ranking Approach to Stress Prediction for Letter-tosyllabification of italian, A generic tool for the Phoneme Conversion, in: Proceedings of the Joint automatic syllabification of Italian (2014) 73–77. Conference of the 47th Annual Meeting of the ACL [17] C. R. Adsett, Y. Marchand, Are Rule-based Syl- and the 4th International Joint Conference on Natlabification Methods Adequate for Languages with ural Language Processing of the AFNLP: Volume 1 Low Syllabic Complexity? The Case of Italian, in: Volume 1, ACL ’09, Association for Computational P. Wagner, J. Abresch, S. Breuer, W. Hess (Eds.), Linguistics, 2009, p. 118–126.

Sixth ISCA Workshop on Speech Synthesis, Bonn, [26] L. P. Dinu, An approach to syllables via Germany, August 22-24, 2007, ISCA, 2007, pp. 58– some extensions of marcus contextual gram63. mars, Grammars 6 (2003) 1–12. URL: https:// [18] C. R. Adsett, Y. Marchand, V. Keselj, Syllabifi- doi.org/10.1023/A:1024089129146. doi:10.1023/A: cation rules versus data-driven methods in a lan- 1024089129146. guage with low syllabic complexity: The case of [27] L. P. Dinu, A. Dinu, On the data base of romanian italian, Comput. Speech Lang. 23 (2009) 444– syllables and some of its quantitative and cryp

References

[1] S. Suyanto , Incorporating syllabification points into a model of grapheme-to-phoneme conversion , International Journal of Speech Technology 22 ( 2019 ) 459 - 470 .

[2] V. N. Vitale , F. Cutugno , A. Origlia , G. Coro, Exploring emergent syllables in end-toend automatic speech recognizers through model explainability technique , Neural Comput. Appl . 36 ( 2024 ) 6875 - 6901 . URL: https://doi.org/10.1007/s00521-024-09435-1. doi: 10 .1007/S00521-024-09435-1.

[3] A. Dinu , L. P. Dinu , On the syllabic similarities of romance languages , in: A. F. Gelbukh (Ed.), Computational Linguistics and Intelligent Text Processing , 6th International Conference, CICLing 2005 , Mexico City , Mexico, February 13-19 , 2005 , Proceedings, volume 3406 of Lecture Notes tographic aspects, in: N. Calzolari , K. Choukri , A. Gangemi , B. Maegaard , J. Mariani , J. Odijk , D. Tapias (Eds.), Proceedings of the Fifth International Conference on Language Resources and Evaluation , LREC 2006 , Genoa, Italy, May 22 -28, 2006 , European Language Resources Association (ELRA), 2006 , pp. 1795 - 1798 .

[28] L. P. Dinu , A. Dinu , On the behavior of romanian syllables related to minimum efort laws , in: Proceedings Workshop Multilingual Resources, Technologies and Evaluation for Central and Eastern European Languages, co-located with RANLP 2009, Borovets , Bulgaria 2006 , 2009 , pp. 9 - 13 .

[29] S. Chebanow , On conformity of language structures within the Indoeuropean family to poisson's law , Comptes rendus de l'Academie de science de l' URSS 55 ( 1947 ) 99 - 102 .

[30] G. Altmann , Prolegomena to Menzerath's Law, Glottometrika 2 ( 1980 ) 1 - 10 .

[31] A. M. Ciobanu , A. Dinu , L. P. Dinu , Predicting romanian stress assignment , in: G. Bouma, Y. Parmentier (Eds.), Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2014, April 26- 30 , 2014 , Gothenburg, Sweden, The Association for Computer Linguistics, 2014 , pp. 64 - 68 . URL: https://doi.org/10.3115/v1/e14- 4013 . doi: 10 .3115/ V1/E14-4013.

[32] L. P. Dinu , A. M. Ciobanu , I. Chitoran , V. Niculae , Using a machine learning model to assess the complexity of stress systems , in: N. Calzolari , K. Choukri , T. Declerck , H. Loftsson , B. Maegaard , J. Mariani , A. Moreno , J. Odijk , S. Piperidis (Eds.), Proceedings of the Ninth International Conference on Language Resources and Evaluation , LREC 2014 , Reykjavik, Iceland, May 26 -31, 2014 , European Language Resources Association (ELRA), 2014 , pp. 331 - 336 . URL: http://www.lrec-conf.org/proceedings/ lrec2014/summaries/1200.html.

[33] K. Cho , B. Van Merriënboer , C. Gulcehre , D. Bahdanau , F. Bougares , H. Schwenk , Y. Bengio , Learning phrase representations using rnn encoderdecoder for statistical machine translation , arXiv preprint arXiv:1406.1078 ( 2014 ).

[34] D. P. Kingma , J. Ba , Adam: A method for stochastic optimization , arXiv preprint arXiv:1412.6980 ( 2014 ).