“Onomatopoeia is the naming of a thing or action by a vocal imitation of the sound associated with it (such as buzz or hiss)” (Britannica, 2024). The phenomenon in question is not restricted to a particular part-of-speech (POS), however due to lexical and functional similarities researchers tend to associate onomatopoeia primarily with interjections, which is seen in some manuscript titles, such as Diccionari d’onomatopeies i altres interjeccions: amb equivalències en anglès, espanyol i francés by Riera-Eures (2010). In fact, sophisticated criteria are needed to distinguish both terms from each other, which is why entire works, even PhD theses, are devoted to this issue (Meinard, 2022).

Interjection, on its turn, is defined as “an exclamatory word or phrase used to express an emotional reaction or to emphasize a thought” (Britannica, 2024). Once compared the definitions of these narrowly interrelated terms, we can conclude that the differences between both phenomena lie in their semantic meaning: while the onomatopoeia expresses sounds, the interjections convey emotions. Rodríguez Guzmán infers a set of additional points of inflection (formal characteristics, syntactic function in sentences) and disjunction (motivation patterns, morphonological processes, semantics) (2011, p. 173) between onomatopoeia and interjection, and finally concluding that both are to be considered as separate word classes (2011, p. 173). If we interpret the word class as part of speech in the context of data mining, particularly, in corpus linguistics, most corpora managers and taggers rely on worldwide conventions, among which the most widespread is the Universal Dependencies (UD) framework, currently used to annotate thousands of corpora. While the UD POS tagset does include interjections, the onomatopoeia does not form its part (Universal Dependencies, 2024). Similarly, many other traditional POS lists, whose number typically ranges between 9 and 10, comprise interjections (since the Latin grammars) but traditionally exclude onomatopoeia. This seems consistent with the logic: while, indeed, onomatopoeia semantically stands alone from other POSs, expressing sounds, the semantic category of a word is not the decisive criterion to assign it a POS label: otherwise, verbal nouns, such as participation or engineering should be semantically classified as verbs.

Both grammar and semantic characteristics come into play when assigning a POS property to a word. From a grammatical standpoint, the onomatopoeias are unchangeable words, as well as the interjections. Furthermore, depending on a particular linguistic school, the sound-imitating meaning is listed among the semantic properties of interjections, which is the traditional approach in Ukrainian grammar (see, for instance, Karamysheva, 2017, p. 218). This feature is also empirically validated by the sampling from the corpus GRAK (see Table 9 and Table 10), where including the interjection tag in the query yields an impressive number of onomatopoeias. In this case, the concept of interjection turns out to be broader than that of the onomatopoeia. Now, the central question is what POS status should apply for onomatopoeia? Beyond the ongoing debates about their part of speech status, there is an immediate need to retrieve sound-imitating words from corpora or another source for various practical purposes. Authors, translators, and editors may need to express sound not only using pure sound imitation but also through derived nouns, verbs, adjectives, and adverbs with the semantics of sounds. Are these words to be classified as onomatopoeias? According to the Merriam-Webster Dictionary, onomatopoeia can also refer to the words formed by onomatopoeia (2024). Thus, not only buzz, hiss and similar words, but also buzzing, hissing, buzzy, and hissy, as well as other onomatopoeically derived lexemes (nouns, verbs, adjectives, adverbs), may form part of this list, as seen (particularly, but not exclusively) in Bidaud’s work (2022), who focuses their research on verbal onomatopoeias. Moreover, in English, assigning a part of speech property for a word such as buzz may be particularly challenging. It is obvious now that the POS-attribution may depend on the semantic and grammar approach elaborated upon in a particular linguistic school, but, from the standpoint of data mining at the current stage, we are to conclude the following:

Given these premises, we qualify hereafter interjectional onomatopoeias as onomatopoeias morphologically characterized as interjections, which act as grammatically unchangeable words and semantically express sounds.

It is important to note that the issue of translating onomatopoeic words comprises two faces: whereas in translation many context-driven techniques allow for multiple ad hoc contextual solutions, for the sake of dictionary compilation, more comprehensive solutions covering a wide range of potential situations in translation are needed.

From the translatological point of view, Yaqubi et al. outline a dozen techniques for using sources to translate onomatopoeia in the following order of precedence:

The mentioned procedures are of enormous practical help for translators who are constrained to bridge numerous onomatopoeic lacunas. Whereas in one language there may be a traditional way of imitating the sound of a particular object or phenomenon, in other languages there might not exist a conventionally accepted onomatopoeia for a given communicative situation, where in yet another, less specific onomatopoeia might be acceptable (i.e., by means of hyperonymic substitution). The onomatopoeia of scissors or other cutting tools in Ukrainian is чик-чик, although no specific sound-imitating word is provided in the list of onomatopoeic words for Spanish, whereas in English and Portuguese there seem to exist analogous onomatopoeias: snip, and rip, as shown in the following examples 1 and 2.

At the same time, a similar sound in Spanish can be traditionally imitated through zas-zas, whose meaning denotes many types of noise, i.e., is a hyperonym (see example 3):

Beyond any possible valuable technique (reproduction, substitution, addition) that may serve as a brilliant situational workaround, it is crucial to explore the existing bilingual and monolingual dictionaries of onomatopoeia first to discover the possible lexicographic gaps and attempt to bridge them, very much in accordance with Yaqubi’s et al. recommendation of looking for an established equivalent as a priority method (2018, p. 220). Nevertheless, borrowing onomatopoeias from the source into the target text looks like a widespread technique. Rodríguez Guzmán presents a list of onomatopoeic words loaned and borrowed into Spanish from other languages (2011, p. 157). At the same time, some tricks used in translation due to the absence of a better suitable equivalence in the target language may indeed be due to a gap or lack of knowledge of existing specific means. This implies that specialized informational resources (such as dictionaries or corpora) would be of great help for translators.

It is obvious that, to compile bilingual or multilingual dictionary entries, credible sources and robust methodology are required. From the standpoint of the 21^th century lexicography, the undisputed number one source for extracting extensive linguistic data are large-volume language corpora, and onomatopoeia is not an exception. In fact, many researchers use their own custom corpora to perform manual searches. For instance, Yaqubi et al. use their research corpus to calculate the frequency of the onomatopoeia in the Charles Dickens novel A Tale of Two Cities and for performing manual searches to retrieve examples of onomatopoeia for their two translations into Persian (2018, p. 211–212).

Some papers do pursue the objective of automating this operation: Orrequia-Barea and Marín-Honor explore techniques particularly focused on onomatopoeic word extraction from large-volume corpora, such as the British National Corpus (2020, p. 47). Nevertheless, works of this kind are few, and our objective is to propose a method to optimize retrieving onomatopoeic words from large corpora and to evaluate their effectiveness in terms of precision involving four languages: English, Spanish, Portuguese, and Ukrainian.

To find rational ways of retrieving examples of onomatopoeia, first we need to find out their most relevant features. As a starting point for observation, we have primarily used the ready-made lists of onomatopoeias in English (Yourdictionary, 2021), Portuguese (Riondlearn, 2022), Spanish (Fundeu, 2011), and Ukrainian (Божко, 2023). These lists were subject to observation with the purpose of extracting some valid markers to be used as reliable formal criteria during automatic or semiautomatic extraction of onomatopoeic words out of corpora. The method of observation, aimed at an intuitive selection of relevant features, yielded the following recurrent (but not exclusive nor mandatory) characteristics of the onomatopoeic words:

The second and fourth observations partially coincide with the patterns proposed by Orrequia-Barea and Marín-Honor (2020, p. 52). The repetition in linguistic sounds’ representation or, specifically, reduplication are known as a universal linguistic feature:

The fifth observation, once implemented in corpus queries, did not produce noteworthy results. Each of the other four observed items merit particular attention and study, and we implemented the detected features in corpus queries. In the current study, our purpose is to focus upon the syllable’s reduplication pattern and the possibilities of its usage to automate onomatopoeia extraction from corpora. Therefore, the methodology is based on the phonic properties of onomatopoeias, such as repeated syllables, and, where possible, upon the part of speech parameters. Since there are no phonically annotated large corpora of the languages in question (Ukrainian, English, Spanish, and Portuguese), we were constrained to base our queries on grapheme levels instead of sound or phonemes.

To achieve the results, we use particular query languages depending on the search engine the corpora are provided with the Corpus Query Language, CQL (Sketch Engine, n.d.) or Corpus Query Processor, CQP (Evert S. & The CWB Development Team, 2022) because they allow for searching patterns matching both regular expressions and linguistic annotation tags.

The automation of onomatopoeia retrieval as an idea started being explored in 2020. Although our interest in the subject appeared independently from the existing studies using similar techniques, we retrospectively took into account all the achievements that are very much in accordance with the current paper.

Orrequia-Barea and Marín-Honor (2020) systematize different graphic properties of onomatopoeic words in written text to extract them by using the regular expression syntax, with some interesting observations regarding the correlation between onomatopoeic formal pattern and the ontological nature of the represented sound using Round & Kwon’s concept of phonaesthemes, i.e. “recurrent pairings of sound and meaning” (Round & Kwon, 2015, p. 2).

Orrequia-Barea and Marín-Honor searched the text of comics to retrieve onomatopoeia in Spanish and French, and for texts in English the regular expression syntax was applied:

More observation on potentially universal features of onomatopoeias comprising some sound combinations are described by Assaneo, Nichols & Trevisan:

Some researchers utilized dictionaries as a starting point for retrieving the onomatopoeic words out of lexicographic sources explicitly marked as onomatopoeic:

A similar approach was implemented by Yaqubi et al. (2018, p. 212) who compiled a list of Persian onomatopoeias at the initial stage of their research. Although this methodology seems promising, we did not apply it in the current study, for it implies significant manual work and does not allow for extraction automation. Moreover, at the current stage, where numerous onomatopoeic lemmas or their graphic variants are not yet included in dictionaries, we chose to extract sound-imitating interjection from corpora including occasional ones. This is why we opted for patterns based on syllables or grapheme repetitions.

We are aware that the patterns based upon repetitions ignore single-syllable words. However, the basic idea relies upon the fact that the repetitions may serve as an access point to further retrieve another kind of onomatopoeia that is not based upon repetitions, some of which are present in the nearest context. In other words, if an onomatopoeia exists in the form of repeated syllables, then it is likely to appear in its monosyllabic or isolated variant, as seen in the following examples (4, 5 and 6) in Portuguese and Ukrainian:

After elaborating upon the enquiry method, we needed to choose the best fitting corpora for extracting empirical data among the available corpora provided with the possibility of looking patterns of texts matching regular expressions.

To subject to test our hypothesis by means of corpus query patterns that we further propose, we need to decide what corpora will serve as the source of empirical basis, which is why we resort to corpora of four languages in which we can read and, consequently, carry out contextual analysis: English, Portuguese, Spanish, and Ukrainian. The number of languages used is also constrained by the accessibility of specific query languages in the corpora interfaces, suitable with the proposed patterns. Although English and Ukrainian are genetically distant from Spanish and Portuguese, the conclusions drawn on four languages from different groups will allow for better judging over the validity of the query patterns. An additional reason to choose English, Romance languages, and Ukrainian was the fact that, for the latter, the bilingual lexicographic contributions in the field of onomatopoeic dictionaries is particularly fruitful, as seen in 2.2., whereas English-Ukrainian, Portuguese-Ukrainian, and Spanish-Ukrainian language combinations are not provided with lexicographic sources containing either interjections or onomatopoeias. Moreover, among the four mentioned languages, only the Ukrainian corpus is provided with correctly tagged interjections, which helps bring to light additional properties of the queries performed.

The best intuitive choice of corpora may seem the referential standard since the reference corpora better represent the general properties of a language. At the same time, with regard to the needs of the research, we are also constrained by the corpora technical details. Orrequia-Barea and Marín-Honor stressed on the downloadability of the corpora:

To overcome this difficulty, we resorted to corpora provided with the CQL (Sketch Engine, n.d.) and CQL query language (Evert S. & The CWB Development Team, 2022), whose usage is illustrated in the next section, which allows for the usage of regular expressions; therefore, it was not mandatory to download any corpus.

While the Portuguese reference corpus CRPC, Corpus de Referência do Português Contemporâneo (CLUL, Centro de Linguística da Universidade de Lisboa, 2008–2016) is provided with the CQP query language search engine, the reference corpora of English are focused on particular countries. The referential corpus of Spanish CREA (Real Academia Española, n.d.) does not allow for the usage of CQL or CQP and, logically, search by means of regular expression, and there is not any referential corpus for Ukrainian. However, given the fact that the team of the corpus GRAK team is making efforts to meet the referential criteria of the corpus and considering another advantage that interjections in this corpus are correctly tagged, we judge the corpus GRAK as the best source for achieving the goal set.

On the other hand, given the fact that onomatopoeias, especially occasionally created words, are likely to appear not only in fiction, but also in internet communication, we finally decided to use available CQL or CQL based referential of internet corpora of English, Portuguese, Spanish, and Ukrainian, particularly:

Corpus Query Language (CQL Guide), Corpus Query Processor (CQP, 2022) or alternative similar querying methods allow searching for given patterns based upon sequences of letters employing regular expression and corpus annotations.

Since many occasional onomatopoeias are not lemmatized nor annotated with particular tags in the corpus, their examples are to be chosen by the attribute word, which is aimed at selecting tokens with specific character sequences independently from their lexical and grammar properties, as shown in the example of Query 1:

The snippets inside the quotation marks composing the major part of Query 1 in CQL and CQP are processed as regular expressions. The regular expression between quotation marks is designed to match words that follow a consonant-vowel-consonant structure, where the first and last consonant combinations in syllables are the same, while the vowels may be either identical or different, e.g., tic-tac, brum-brum, etc. Let us now provide a detailed explanation regarding each part of the regular expression used in Query 1 in Table 1:

Finally, the entire expression matches the pattern of at least two consonant-vowel-consonant hyphenated similar syllables with the same consonants and varying vowels. More detailed explanations of the CQL syntax usage are accessible in the Corpus Query Language Guide (Sketch Engine, n.d.).

Query 1 is well suited for the corpora of English. However, in the case of applying a similar query to a language with different character sets (diacritics, Cyrillic, Greek, etc.), the characters inside the regular expressions are to be adapted to its alphabet, as we do for Portuguese, Spanish, and Ukrainian. To apply the same query for the Corpus de Referência do Português Contemporâneo (CLUL, Centro de Linguística da Universidade de Lisboa., 2008–2016), Portuguese Language Corpus from the Leeds Collection of Internet Corpora (2022), we need to extend the character class with the diacritics (Query 2). Query 3 is respectively adapted for Spanish, and Query 4 for Ukrainian:

We consider valid those examples that are interjectional onomatopoeias, i.e., sound-imitating words belonging to the class of interjections. Since the corpus query cannot delimit the sound-imitating words from those phonically motivated lexemes that are no longer interjections, but that could be qualified as such in the moment of creation, we considered those cases of “etymological” onomatopoeias as valid examples (e.g.: zigzag, criss-cross, flip-flops, etc.). Although this decision may seem arbitrary, we aim to evaluate the queries’ potentiality to match the necessary graphic patterns, rather than exploring their usability for distinguishing the evolution of the word meaning. At the same time, we discard from this survey other types of onomatopoeia expressed with nouns and verbs. For instance, the Query 7 (Table 2), among results, yields the words murmur and barber, that might be valid for other objectives. Whereas conventional onomatopoeias could be checked in the dictionaries, to judge the onomatopoeic function of occasional words, we use contextual analysis at the level of concordance line or paragraph.

The representation of repeated syllables through grapheme level posits a series of questions such as syllables division, diphthongization and hiatus, unpronounced graphemes, and open and closed syllables. Nevertheless, some of these dilemmas can be overcome by assuming that onomatopoeic words are not created solely according to strict phonic patterns: sometimes the repetition may include one or several syllables (meow, meow-meow) and some onomatopoeias may present variants of graphic representations (achoo, atchoo, achew); therefore, there is probably little sense of rigorous compliance to the syllable divisions. Additional observations can shed light upon the fact that the vast majority of the onomatopoeias start with an initial consonant grapheme, end with another consonant grapheme, and, in separate cases, with a vowel. In other words, the main relevant feature to take into account for the corpus queries is to consider repetitions of sequence starting with consonant graphemes, followed by vowels and optionally ending with another consonant grapheme or a group of such graphemes.

Given that the hyphen can also be optional, the possible queries are to rely upon a four-member paradigm:

Hereafter, in Tables 2, 3, 4 and 5 we expose the results yielded by the respective queries with the valid examples out of 100 first generated lines in the concordances; we indicate the number of repeated forms in parenthesis.

According to the results obtained, an outstanding fact that drew our attention was that for the queries 5, 6, 9, 10, 13, 14, 17 and 18, the rate of valid examples was much higher than those yielded by others. Many cases of onomatopoeias do not figure in the reference explanatory dictionaries. For example, out of 25 different onomatopoeias retrieved from the CRPC by queries 9 and 10, 17 do not appear as entries in the dictionary Priberam (2023) neither in bisyllabic or monosyllabic forms as sound-imitating lexemes: cri-cri, frou-frou, wha-wha, glu-glu, tai-tai, chi-chi, flic-flac, can-can, tan-tan, tim-tim, flics-flacs, zig-zag, zuns-zuns, den-den, hip-hop, tic-tac, tam-tam. This means that significant parts of the examples found are occasional sound-imitating words missing in lexicographic entries of modern dictionaries.

The impressive number of retrieved examples of repeated non-hyphenated open syllables observed in Portuguese and Ukrainian is likely due to a universal linguistic tendency. In the Leeds Corpora, the interface limited the maximum number of examples to 1,000, which is why we could not access the complete results.

From Tables 2, 3, 4, 5 we can also observe that the more overall concordance lines are generated as per the query, the less specific the query focus of the onomatopoeias is.

Table 6 integrates the number of specific results produced by the queries 5–20 in the four corpora excluding the types of queries that yielded 0 results.

It is obvious that the repeated hyphenated closed syllables patterns yielded the most significant results among the four languages. To confirm that this is a tendency rather than an occasional combination of factors, we transposed the data as in Table 7 and subjected it to an ANOVA test (single-factor), as we expected that in this experiment the only significant factor was the type of syllables.

Table 8 illustrates the results of the ANOVA test performed in Microsoft Excel.

It is seen that the p-value (i.e., the probability that the achieved results are due to random coincidence) is equal to 0.000251. This value is far lower than the conventional 0.05 (i.e., the 5% threshold), which confirms that the data obtained are not due to chance, and the closed-syllable base onomatopoeias with repeated sounds turn out to be the most productive query pattern in the four observed languages, making the hyphen a robust onomatopoeic marker.

It is obvious that most hyphenated onomatopoeias do exist both in bisyllabic and monosyllabic forms, e.g., bang-bang / bang, cling-cling / cling, ching-ching / ching, beep-beep / beep, plink-plink /plink, and this property could be used at a further stage to optimize the extraction programmatically according to the following algorithm: if a pattern with repeated syllables recurrently occurs in a corpus, perform monosyllabic search for the syllabus used in the pattern.

Since interjections in the corpus GRAK have been correctly tagged, there is the possibility to perform the queries now with an additional pos-filter to limit the sampling exclusively to the cases of interjections, which increased the precision approximately twice. The queries 17 and 18 (Table 5) are now extended with this pos-filter. The results are shown in Tables 9 and 10, respectively.

Among the corpora involved in this survey, the only corpus with correctly tagged interjections is the corpus GRAK, hence we put to test only the selection in Ukrainian. The yielded results, once added the condition ([tag=“.*intj.*”]), are shown in Table 9 and Table 10.

In fact, in many corpora the interjections appear tagged as nouns or adjectives. A serious challenge of modern corpus linguistics is the interjections recognition in transcribed corpora (Tellier et al., 2010):

The results of Query 22 (Table 10) confirm the demand for text corpora containing accurately tagged interjections. What catches our attention in Table 9 and Table 10 is the significantly superior performance of the query in the corpus GRAK, which is seemingly due to the presence of correct interjection annotation, whereas in the corpora of English, Spanish, and Portuguese utilized, the interjectional onomatopoeias are assigned tags of nouns or adjectives. This pos discrepancy makes it impossible to use the pos-attribute as a reliable feature for extraction.

The logical question arising from the observations of Table 9 and Table 10 is whether the query takes into account any linguistic feature. The corpora are textual databases provided with linguistic relevant data, but our queries performed to all the corpora except the corpus GRAK were based exclusively upon formal features such as grapheme combinations, with no relation to other linguistic properties. In fact, most queries include the attribute word. From the standpoint of corpus linguistics, word is a specific sequence of symbols separated by delimiters such as space. This means that the queries are intended to search particular sequences of letters by the regular expression syntax, which can also be successfully searched in many text editors with regular-expression engines, such as Notepad++, Sublime edit, or similar programs by applying the following query:

It can be observed that the regular expression from Query 23 was the same as in Query 22 (Table 10). As expected, the search within the novel La sombra del águila (‘The Eagle’s Shadow’) (Pérez Reverte, 1993) yielded 100% of the valid results: cling-clang (seven times), bang-bang (three times), zas-zas (two times), and ras-ras (once). This result illustrates an ideal accuracy, but not in term of precision, since some valid cases of onomatopoeia might have been left out by the query. This text is rich in onomatopoeias since this novel narrates war events, which was the main reason to use it as additional empirical data.

The regular expression from the query used in Table 11, once applied to the text by Arturo Pérez Reverte, in contrast, did not yield purely onomatopoeic examples, however the found matches were related to another sound imitating phenomena, since they all imitated stuttering speaking of a character: vu-vuelto, lo-locos, lo-locos, su-suicidio, la-la, va-van, de-descuartizar, te-temo, po-posible, ma-malentendido, la-lapsus, he-hemos, po-polvo, ci-ciento, ma-mañana, su-suman, co-compañía, lui-la, pa-parece, se-setecientos, he-heridos, ci-cierto, du-duele, so-sombra, sa-sacrificio, ge-gesta, pe-perdido, nu-nuevecitas, de-demás, po-podéis, mi-mierda (Pérez Reverte, 1993).

Given the constraints of available space, we are unable to explore the marker based upon letter repletion in greater detail, but it is worth outlining some preliminary observations to be put to the test in further research. The consonant repetition at least three times as an onomatopoeic marker was earlier observed by Orrequia-Barea and Marín-Honor (2020, p. 53). Our observations suggest that additional markers’ usage in queries to transcribed text corpora, such as exclamation marks, ellipses, or quotation marks, may significantly improve the results. We should also admit that both vowels and consonants are relevant in this pattern. The results are shown in Table 11.

The sound sources are easily deducible from the nearest context, as shown in the examples 7, 8 and 9:

Many onomatopoeic words are known to be neologisms, nonce or occasional words due to their creative nature. Therefore, they must appear in modern corpora as non-lemmatized wordforms, i.e., they must be stored in databases as words under unknown or empty lemmas. Hence, in the process of automatic lemmatization, the occasional onomatopoeias are not recognized as lexemes belonging to the given vocabulary and are assigned empty or “unknown” lemmas. CQL and CQP are provided with the possibility of searching for this kind of lemmas. In other words, rare lemmas, searched through the queries [lemma=””], [lemma=”\|\|”] or [lemma=”unknown”] depending on conventions of a given corpus can increase the chance of finding exotic or occasional onomatopoeias.

Another promising formal feature to delve deeper into is the frequency factor, which can be also utilized as a marker: The transcribed onomatopoeias are likely to show lower frequency in the corpus in comparison to commonly used words, and Sketch Engine corpora provided with the CQL allow for applying the frequency as a separate filter in the query. The aforementioned and possibly other markers seem to be a promising perspective for research in further surveys.