English. In recent times, linguistic research and computational linguistics have considered the morpho-semantic properties of the grammatical number in nouns. In this context, our study aimed at using a language model to test number variation on selected Russian somatic idioms containing locally and generally marked nouns. Overall, the model's sensitivity corroborated the relevance of number in idioms, especially in instances where the semantics of number is particularly significant. Italiano. Negli ultimi tempi la ricerca linguistica e la linguistica computazionale hanno iniziato a considerare le proprietà morfo-semantiche del numero grammaticale dei sostantivi. In tale contesto, questo studio si è proposto di utilizzare un modello linguistico per testare la variabilità del numero nei fraseologismi somatici russi con nomi dalla marcaztezza locale o generale. Nel complesso, la sensibilità del modello al cambiamento di valori grammaticali ha dimostrato l'influenza del numero, soprattutto in istanze in cui questo è particolarmente significativo dal punto di vista semantico.
Somatic idioms are defined as phraseological units which contain lexemes referring to human or animal body parts [1, 2]. This class is found in many languages, due to the universal nature of the somatic lexicon. Somatisms are usually observed from a semantic perspective and mostly in contrastive analyses, since their meaning is regarded as a cultural-specific conceptualisation. However, to the best of our knowledge, existing inquiries, both in theoretical and computational linguistics, have not yet considered the morpho-semantic features of the grammatical number [3] as part of the overall idiomatic meaning. Furthermore, our interest in number is motivated by the fact that many somatic terms exhibit an efect of the pluralization’s lexical properties. In particular, nouns denoting bipartite or complex body parts, which are naturally constituted by a plurality, show a higher frequency in the plural form in many languages. This condition is defined local or semantic markedness [4, 5], as the singular value becomes the locally marked member of the opposition, changing the morpho-semantic dynamics between the two values. The goal of our work is to consider how change in the grammatical number afects the idiomatic structure of somatisms. More specifically, the questions underlying the following experiment were:
on the idiomatic structure of Russian somatic idioms? 2. Does the type of markedness afect the probability of finding a singular or a plural value?
an efect on the overall structure of somatic idioms, especially in the cases of local markedness. In detail, we assumed that if we switched the values of a locally marked somatic term a greater impact on the idioms would be observed, compared to generally marked somatic terms.
The following sections describe how we addressed these questions, starting from an examination of related works. The new dataset of Russian somatic idioms that we constructed to test our hypotheses and its creation method are thereafter illustrated1. Subsequently, our experiment is presented, which was conducted with ruBERT, a Russian-trained BERT, following a methodology similar to Salazar et al. [6] for acceptability, and Pedinotti et al. [7] for semantic plausibility. The results of the experiment are ofered and discussed in the final sections. Ultimately, a conclusive section provides an answer to our questions, highlighting the relevance of our work and possible future expansion.
corresponding contexts, showing how diferent meanings associate with the numerical values. On the contrary, Studies on idiom semantics adopting a compositional мозг - mozg “brain" exhibits a quite unique condition: alapproach have claimed that the semantic analyzability though it denotes a single item, its pluralization features or decomposability of idioms (the meaning of the con- an additional meaning, absent in the singular, acting as a stituents contributes to the overall figurative meaning) is separate lexeme (as does the English equivalent ‘brains’). related to their flexibility, that is their lexical or syntacti- Regarding the markedness, the Academic Russian Gramcal variation (see Cserép, Dobrovol’skij for an overview). mar [15] and Ušakov’s Dictionary [16] indicate that some
In idiom variation research, number has been inves- nouns are mostly used in the plural, including somatic tigated as a flexibility dimension by Langlotz [ 10] and terms. However, a full account for the whole somatic Cserép [11]. Specifically, Langlotz [ 10] asserts that num- lexical domain, in terms of frequency and markedness, is ber variation is systematical in VP-idioms or PP-idioms not available as it is for other languages. that feature an isomorphic, that is analyzable, semantic Only in recent times have the lexical properties of structure. However, this variation is believed to be pre- number been addressed in the computational field. Grovented if the idiom contains an invariable noun or a noun mann and Declerck [17] investigate the semantic shifts which is incompatible with the global idiomatic meaning. created by the regular inflectional morphemes of number. The second study [11] adopts a corpus-driven approach The study shows the impact of these inflectional variants and it retrieves the morpho-syntactic alternants in the on morphological embeddings and their variability, on a noun phrase of V NP-idioms. As a result, a higher fre- par with derivational morphemes. Even more recently, quency in the singular value is assessed and a similar Janzso [18] focuses on the ambiguity originating from correlation between number variation and decomposabil- these inflectional variants, as the plural form of some ity is found. Both works acknowledge the role of number nouns carries a secondary non-grammatical meaning, abin the idiomatic structure, however they are limited to sent in the singular value. Specifically, the work inquires the English language and no distinction based on the lex- about how ambiguity in number and gender is treated ical meaning or type of markedness of the constituents by contextual meaning representations, using four prehas been drawn. trained BERT models on the disambiguation task. In the
In concordance with the aforementioned theories on literature on idioms, computational studies focused on the category of number, studies on the Russian language their identification or comprehension [ 19, 20, 21, 22], but assert that «the grammatical values of number are closely to the best of our knowledge there have been no studies related and actively interact with lexical meaning» 2 [12, investigating the role of number in the construction of 24] (see also Vinogradov [13]). As observed by Ljaevskaja idioms and its interpretation. [14], this interaction is particularly evident in nouns that denote complex items, including complex and paired body parts. The lexical meaning of such nouns is inter- 3. Dataset twined with their numerical information as they denote a complex structure, constituted by plural members func- The dataset was constructed by selecting somatic idioms tioning or perceived as a whole. As a result, in these cases from two Russian phraseological dictionaries [23, 24]. the grammemes of number convey more than just the Nine somatic components were considered in the selecnumerical information. To assess this extra-grammatical tion: six terms referring to bipartite or complex body meaning, the two values of some somatic terms have been parts (глаз - glaz “eye", рука - ruka “hand/arm", нога observed, based on their use in context [14, 71-86]. The noga “foot/leg", губа - guba “lip", ухо - ucho “ear", зуб result of this detailed qualitative investigation is the dis- zub “tooth"); three terms referring to a single body part tinction of two contexts: regular distributive contexts (i.e. or organ (голова - golova “head", мозг - mozg “brain", the standard grammatical opposition: one - many); non- язык - jazyk “tongue"). distributive contexts, which are mostly noun-specific, In order to have a broad idea of the dominance convalue-specific, and language-specific (As Ljaševskaja il- dition [5], and therefore of the type of markedness, the lustrates [14, 72], он шел по колено [sg] в грязи - on relative frequency for each term was obtained from the sˇel po koleno v grjazi cannot be replaced with the plural Russian National Corpus3 (see Table 1). он шел *по коленам [pl] в грязи - on sˇel po kolenam v These components were considered both in the singrjazi, whereas in English a plural form is required: “He gular and in the plural form. Altogether 73 somatic walked in mud up to his knees"). Furthermore, similar idioms in the singular and 73 somatic idioms in the contexts lead to a classification of similar somatic terms. plural were retrieved4. Idioms registered in both forms Apparently, глаз - glaz “eye" and ухо - ucho “ear" display
и активно взаимодействуют с лексически-ми значениями».
4Ideally, ten idioms were to be selected for each term. However, the actual number of idioms is due to their availability in the dicour dataset consists of 438 sentences, divided into three diferent conditions: Original (73 with singular form and 73 with plural form), Manipulated and Control sentences as in the following table 26.
Noun глаз - glaz “eye" рука - ruka “hand/arm" нога - noga “foot/leg" губа - guba “lip" ухо - ucho “ear" зуб - zub “tooth" мозг - mozg “brain" голова - golova “head" язык - jazyk “tongue"
Model For our experiment, we used the ruBERT base model (12-layer, 768-hidden, 12-heads, 178 parameters) provided by the Sberbank group (https:// huggingface.co/ai-forever/ruBert-base). This model was built by taking BERT [27] as a basis. ruBERT-base has a Byte-Pair Encoding (BPE) tokenizer with a dictionary of 120 thousand tokens. It was trained with 30 GB of were avoided whenever possible (по губе/по губам - po Russian text, which includes Wikipedia, news, part of gube/po gubam “to one’s taste", lit. “to one’s lip/lips"). the Taiga corpus [28], and some books7. Formally, the selected instances range from a higher degree of idiomaticity (fixed idiomatic expressions as глаза Experiment In order to assess the sensitivity of the на лоб полезли - glaza na lob polezli “eyes nearly popped model to number variation in the selected idioms, we out", lit. “eyes climbed on the forehead") to a more flexible used the pseudo-log-likelihood (PLL) score [31]. As condition (idiomatic prepositional phrases as в головах - shown in Salazar et al. [6], the PLL can be considered v golovach “at the head of", lit. “in the heads"). as a measure comparable to probability. The authors
The Russian National Corpus was queried and the con- demonstrate that the PLL outperforms scores obtained texts were derived and trimmed to simple sentences. To with auto-regressive models in a series of tasks related evaluate the impact of change in number, the original to sentence acceptability. Furthermore, in the work of sentences were manipulated in number (manipulated Pedinotti et al. [7] the same measure is used to assess the condition). As a result, the somatic component in the diference in plausibility between metaphorical sentences, original condition was inflected to the alternative nu- literal sentences and nonsense sentences. The results merical value in the grammatical opposition (i.e. the show a quite good correlation with human judgments of component in the original singular-only form was in- semantic plausibility. lfected in the plural form and vice versa). Subsequently, The PLL of a sentence can be derived by iteraa control condition was created by substituting the so- tively masking individual tokens, one at a time, using the matic component with an alternative word. However, it main function of the MLM (Masked-Language Modeling) should be noted that a random substitution was excluded, model. For each masked token , the probability is calconsidering the aforementioned gradient of idiomaticity, culated based on all other words in the context, and the which relates to an equivalent gradient of substitutabil- log-probabilities for all tokens are summed. This process ity of the idiom’s component parts [25]. Therefore, the is illustrated by the following Equation 1: control condition was created by considering alternative candidates for a given construction ofered by the corpus | | 5. From a grammatical perspective, the original numeri- ( ) = ∑︁ log ( | ∖) (1) cal value was always maintained in the alternative words. =1 On the other hand, the grammatical categories of gender and animacy were maintained as far as possible, consid- 5. Results ering their subjection to the choice of a valid semantic alternative in terms of context (e.g. the original в гла- We calculated the significance of the diference between зах - v glazach [m] “in the eyes" was substituted with the diferent conditions using the Wilcox test for multiple в комнатах - v komnatach [f] “in the rooms"). Finally, comparisons with Bonferroni correction. tionaries or in the Russian National Corpus. Specifically, in the singular section, ухо - ucho “ear" and зуб - zub “tooth" (9 idioms), губа - guba “lip" (2 idioms) and мозг - mozg “brain" (3 idioms) 6Unlike the control and the manipulated sentences, the translation did not reach the set number. The same applies for губы - guby of the original Russian idioms refers to the correspondent English “lips" (6 idioms), головы - golovy “heads" (5 idioms), языки - jazyki idioms given by a Russian-English phraseological dictionary [26] “tongues" (2 idioms) in the plural section. whenever possible. 5In limited cases, mostly in idioms originally in the singular form, a 7We selected this model over DeepPavlov’s RuBERT [29] because it higher degree of idiomaticity rendered necessary a random word. performed better in the RussianSuperGLUE evaluation [30].
−2 s d o o h liek −4 li g o l dou −6 e s p e s ab− −8 T R E B ru −10
Sentence мне было очень стыдно за свой язык без костей - mne bylo ocˇen’ stydno za svoj jazyk bez kostej “I was very ashamed of my loose tongue", lit. “I was very ashamed of my tongue without bones" мне было очень стыдно за свои языки без костей - mne bylo ocˇen’ stydno za svoi jazyki bez kostej “I was very ashamed of my tongues without bones" мне было очень стыдно за свое тело без костей - mne bylo ocˇen’ stydno za svoe telo bez kostej “I was very ashamed of my body without bones" но все это для отвода глаз - no vse e˙to dlja otvoda galz “but all this is for a distraction", lit. “but all this is for the withdrawal of the eyes" но все это для отвода глаза - no vse e˙to dlja otvoda glaza “but all this is for the withdrawal of the eye" но все это для отвода вод - no vce e˙to dlja otvoda vod “but all this is for the withdrawal of the waters"
Plural With regard to idioms containing somatic terms As a matter of fact, the singular form proves to have a in the original plural condition, the diferences between probability that is not significantly diferent from any the original condition and the manipulated condi- other word occurring in the construction. tion, and between the original and the control condition are significant ( − < 0.005). On the other Singular As for the singular, we can observe a diferhand, there is no significant diference between the use ence between the original condition and the other two of the singular (manipulated condition) and the difer- (see Figure 2), but there is no significance from the staent word in the control condition (we can also observe tistical analysis we subjected the data to. The lack of the diferences in the Figure 1). This may be evidence significance could be due to a lower frequency of idioms of a real distance between the plural and singular forms. featuring somatic parts in the singular form. Nonethes d o o ih −3 l e k i l g o l o du −6 e s p e s a b − T−9 R E B u r −12
seey scee_y ssee_y sandh scad_nh ssad_nh tfee tfece_ tfsee_ lisp ilsc_p ilss_p tteeh ttceeh_ ttseeh_ rsae rscae_ rssae_ saedh scaed_h ssaed_h tseguno tsecu_gno tsseu_gno irsanb irscan_b irssan_b less, the data show interesting trends when analysing the grammatical number in the terms considered, specifibehaviour of individual nouns more in detail (as we can cally to their type of markedness. Firstly, it could be see in Figure 4). plausible that the higher significance of the plural section correlates with local markedness. As table 1 shows, most terms (six out of nine) display a condi6. Discussion tion of plural dominance, therefore the plural form is the naturally expected value. Secondly, in both sections As it is evident from the results, we can positively address these six locally marked terms show a diference between the first question, asserting that the grammatical num- the original condition and the manipulated condiber of nouns significantly afects the structure of so- tion, demonstrating the impact of number. On the other matic idioms. Indeed, the diferent probability obtained hand, generally marked terms (“head", “tongue", “brain") from the conditions indicate that the two grammatical display little or no diference (see Figures 3 and 4). values are not interchangeable as one would imagine. As Given the influence of the individual terms on both a result, an overall morphological flexibility of the idiom sections, it is worth discussing them separately. associated with number variation as intended in [11] has not emerged.
The diference between the two values is significant Locally marked terms “Eye" and “ear" occupy a dedin the plural section (see Figure 2), as it shows a higher icated section in Ljaševskaja [14]’s investigation, as it probability to find a plural form in idioms originally in was observed that they have similar contexts. This simithe plural (see Figure 1). On the other hand, the singular larity seems to be confirmed both in the plural and in section displays a weak diference, except in two cases the singular section (see figure 3 and 4). However, at a (“lip", “tooth"), whose values result to be less interchange- closer look, “ear" behaves similarly to “hand/arm" and able. “foot/leg", especially in the plural section. These nouns
The diferences between and within the two sections present a significant diference between the original could be due to the morpho-semantic properties of the plural condition and the manipulated condition (in
aehd eacdh_ eadh_p anhd acdnh_ adnh_p tegnou tecguno_ teungo_p tfoo tfcoo_ tfoo_p eey eecy_ eey_p rae reac_ rea_p irabn iracnb_ iranbp_ ilp lic_p li_pp ttooh ttcoho_ tthoo_p “ears" and “hands" the − < 0.005; while in "feet" the singular section do not explicitly confirm the existhe − < 0.02). tence of the additional sense associated with the plural
An unexpected outcome is ofered by “lip", there- of “brain" (as noted in Ljaševskaja [14, 86]). Nonetheless, fore it cannot be associated with the nouns denoting the semantic discrepancy between the singular and the a bipartite item, despite its referential meaning. More- plural meaning of “brain" could still explain the less likely over, it behaves similarly to a noun denoting a complex plural value in the manipulated condition. body part: “tooth" (which is significant in both sections, − < 0.005). In the singular section “lip" is more probable in the singular value and it could be due to a 7. Conclusions more fixed structure of the original idiom.
that our initial assumptions prove to be partially justiGenerally marked terms Among the generally ifed. Comparing original sentences to manipulated and marked terms “head" and “tongue" seem to follow the control sentences, we have found that, overall, number tendency observed in [11]. In fact, a certain interchange- significantly afects the probability which connects the ability between the values is demonstrated by the sin- constituents of the idiom. Specifically, idiom variation apgular section; while in the plural section, the original pears to be subject to the morpho-semantic properties of plural form is more probable, possibly denoting a more the nominal constituent in the idioms considered. Indeed, ifxed structure. our data showed that the result varies depending on
Despite being a generally marked term, “brain", pro- the type of markedness. As a result, change in number vides an unexpected outcome. In fact, it features an evi- for the generally marked constituents is more probable. dent diference between the values in the singular sec- This could confirm the general tendency of number betion, possibly alluding to a more fixed structure in the ing a flexibility dimension in idioms, which does not original idiom. On the other hand, little change is ob- alter the idiomatic structure. On the other hand, locally served in the plural section. The results obtained in marked terms show a significant diference in probability between the two values. The diference suggests a Metonymy and Conceptual Blending, Walter de less probable number variation, which may allude to a Gruyter, 2011. less flexible idiomatic structure. However, some terms [10] A. Langlotz, Idiomatic Creativity: A cognitiveexhibit a peculiar condition, whose interpretation may linguistic model of idiom-representation and idiombe subjected to the formulation of future hypotheses. variation in English, John Benjamins Publishing, 2006.
Limitations and future research Despite the corre- [11] A. Cserép, Idiom variation and decomposability lation between human judgment and the PLL measure part ii: Variation in the noun phrase, Yearbook of shown in Pedinotti et al. [7], the inclusion of human Phraseology (2017). evaluations could still be necessary given the figurative [12] V. Degtjarev, Kategorija čisla v slavjanskich meaning of idioms. This comparison could improve the jazykach. Istoriko-semantičeskoe issledovanie, Izrobustness of the analysis while clarifying the results ob- datel’stvo Južnogo Federal’nogo Universiteta, 2014. tained by the model. Furthermore, given the unexpected [13] V. V. Vinogradov, Russkij jazyk: grammatičeskoe outcome for some nouns, a larger dataset including more učenie o slove, Vyšaja škola, 1986. idioms per nouns and a wider variety of nouns could be [14] O. N. Ljaševskaja, Semantika russkogo čisla, Jazyki designed. Slavjanskoj Kul’tury, 2004. [15] N. J. Švedova; Nina D. Arutjunova, et al., Russkaja
Grammatika, Nauka, 1980.
References [16] D. Ušakov, Tolkovyj slovar’ russkogo jazyka, sov. encikl., 1935-1940. [1] I. Olza-Moreno, Aspectos de la semántica de las [17] D. G. T. Declerck, Towards the detection and forunidades fraseológicas. La fraseología somática met- mal representation of semantic shifts in inflectional alingüística del español, Ph.D. thesis, Universidad morphology, in: 2nd Conference on Language, Data de Navarra, 2009. and Knowledge (LDK 2019), 2019. [2] F. Vakk, O somatičeskoj fraseologii v sovremennom [18] A. Janzso, Disambiguating grammatical number ėstonskom literaturnom jazyke, Ph.D. thesis, Insti- and gender with bert, in: Proceedings of the Student tut jazyka i literatury Akademii nauk ĖSSR, 1964. Research Workshop Associated with RANLP 2021, [3] P. Acquaviva, Lexical Plurals: a morphosemantic 2021.
approach, Oxford Universtity Press, 2008. [19] Y. Dai, Y. Liu, L. Yang, Y. Fu, An idiom reading com[4] P. M. Tiersma, Local and general markedness, Lan- prehension model based on multi-granularity reaguage (1982). soning and paraphrase expansion, Applied Sciences [5] H. Baayen, C. Burani, R. Schreuder, Efects of se- 13 (2023). URL: https://www.mdpi.com/2076-3417/ mantic markedness in the processing of regular 13/9/5777. doi:10.3390/app13095777. nominal singulars and plurals in italian, Yearbook [20] V. Nedumpozhimana, F. Klubička, J. D. Kelleher, of Morphology (1997). Shapley idioms: Analysing bert sentence em[6] J. Salazar, D. Liang, T. Q. Nguyen, K. Kirch- beddings for general idiom token identification, hof, Masked Language Model Scoring, in: Frontiers in Artificial Intelligence 5 (2022). URL: Proceedings of the 58th Annual Meeting of https://www.frontiersin.org/articles/10.3389/frai. the Association for Computational Linguistics, 2022.813967. doi:10.3389/frai.2022.813967. 2020, p. 2699–2712. URL: https://www.aclweb.org/ [21] Z. Zeng, S. Bhat, Idiomatic expression identificaanthology/2020.acl-main.240.pdf . tion using semantic compatibility, Transactions [7] P. Pedinotti, E. Di Palma, L. Cerini, A. Lenci, A of the Association for Computational Linguistics howling success or a working sea? testing what 9 (2021) 1546–1562. URL: https://aclanthology.org/ BERT knows about metaphors, in: Proceedings of 2021.tacl-1.92. doi:10.1162/tacl_a_00442. the Fourth BlackboxNLP Workshop on Analyzing [22] J. Peng, A. Feldman, Automatic idiom recognition and Interpreting Neural Networks for NLP, Asso- with word embeddings, in: Symposium on Inforciation for Computational Linguistics, Punta Cana, mation Management and Big Data, 2015. Dominican Republic, 2021, pp. 192–204. URL: https: [23] A. N. B. D. O. Dobrovol’skij, Akademičeskij slovar’ //aclanthology.org/2021.blackboxnlp-1.13. doi:10. russkoj frazeologii, Leksrus, 2020. 18653/v1/2021.blackboxnlp-1.13. [24] A. I. Fedorov, Frazeologičeskij slovar’ russkogo lit[8] A. Cserép, Idiom variation and decomposability eraturnogo jazyka, Astrel’, AST, 2008. part i: Verbal variation, Yearbook of Phraseology [25] W. O’Grady, The syntax of idioms, Natural Lan(2017). guage Linguistic Theory (1998). [9] D. Dobrovol’skij, Windows to the Mind: Metaphor, [26] S. Lubenskaja, Bol’šoj russko-anglijskij frazeologičeskij slovar’, AST-press, 2004. [27] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova,
Bert: Pre-training of deep bidirectional transformers for language understanding, 2019.
arXiv:1810.04805. [28] S. O. Shavrina Tatiana, To the methodology of corpus construction for machine learning: «taiga» syntax tree corpus and parser., in: Proceeding of "CORPORA2017", international conference, 2017. [29] Y. Kuratov, M. Arkhipov, Adaptation of deep bidirectional multilingual transformers for russian language, 2019. arXiv:1905.07213. [30] T. Shavrina, A. Fenogenova, A. Emelyanov,
D. Shevelev, E. Artemova, V. Malykh, V. Mikhailov, M. Tikhonova, A. Chertok, A. Evlampiev, Russiansuperglue: A russian language understanding evaluation benchmark, arXiv preprint arXiv:2010.15925 (2020). [31] A. Wang, K. Cho, BERT has a Mouth, and It
Must Speak: BERT as a Markov Random Field Language Model, in: Proceedings of the Workshop on Methods for Optimizing and Evaluating Neural Language Generation, 2019, p. 30–36. URL: https://www.aclweb.org/anthology/W19-2304/.