1 Introduction
The Multi3Generation COST Action (CA18231) aims to enhance the importance of Natural Language Generation (NLG) research for societal benefit through collaborative efforts. It advocates inclusivity by embracing valuable resources and methodologies, irrespective of their origin, whether from Large Language Models (LLMs) or linguistically-oriented models. While LLMs have propelled text generation forward, they come with drawbacks that pose significant challenges when used without adequate knowledge and training and need scientific attention. The initial challenge lies in understanding the models due to their opacity, complicating the comprehension of generative processes. Human intervention is limited, resulting in bots prone to hallucination and misinformation. Adapting to various styles and domains also presents difficulties. Additionally, there is concern over the irresponsible use of undisclosed resources, amplifying worries about misuse, especially in domains reliant on trustworthy data processes. Potential risks may exert a more pronounced influence within the critical domains of education, health, law, and industry. In this article, we exemplify controllable linguistic-based approaches (to NLG and Natural Language Processing (NLP) in general) that prioritize the development of resources covering various language features, and characterized by a clearly defined design, distinguished traits, including transparency, adaptability, traceability, and human manageability. An important goal for linguists is to describe natural languages in a manner that enables automatic systems to differentiate between correct sentences, such as The lady ate these cookies from incorrect sequences of words, such as The ate lady cookie these. For a software application to function effectively, it requires access to several key components:
the complete set of linguistic units present in the text, known as the language vocabulary: the, lady, to eat, this, cookie.
the operations that establish connections between the word forms found in the text and the vocabulary, known as language morphology (to eat ⇄ ate, this ⇄ these, cookie ⇄ cookies).
the structure of sentences in the language, referred to as language syntax, enables the system to analyze sentence structures, such as (The lady)Subject (ate)Verb (these cookies)Object.
various distributional constraints among sentence elements, encompassing the semantics of the language, allowing the system to verify compatibility between sentence elements; for example, to eat as a verb requires a human subject and a edible (concrete) object.
However, knowing what constitutes a correct sentence is far from fully describing a language: one also needs to know how to express a negation (The lady did not eat these cookies), how to focus on an argument (It is these cookies that the lady ate), how to pronominalize an argument (She ate them), among others. The important question is: what transformations are applicable to any given sentence? In the realm of NLP application development, the objective is to create software capable of automatically generating any transformed sentence from a given elementary sentence, or conversely, breaking down any complex sentence by producing the elementary sentences containing its predicates, thereby analyzing its meaning.
The rest of the article is organized in the following way: Section 2 presents the procedural approach, which involves describing each transformation with a transducer. Section 3 outlines the declarative approach, where finite state automata describe transformations of the same predicate. Section 4 showcases eSPERTo, a paraphrasing system based on the procedural approach. Finally, Section 5 provides the main insights about both methods and discusses future directions.
Methods
2. The procedural approach
This application can be implemented using various computational devices, including Finite-State Transducers, Recursive Transducers, and Linear-Bounded Automata, such as the one represented by the graph in Figure 1.
Figure 1. Simple transformation.
This graph, which has the power of a Linear-Bounded Automaton, can be automatically applied to any text to perform a three-step computation:
1. it first matches sequences constituted by a pronoun He or She, or a singular determiner (<DET+s>) followed by a singular noun (<N+s>), followed by a verb conjugated in the preterit (<V+PT>), followed by a pronoun them or a plural determiner (<DET+p>) followed by a plural noun (<N+p>).
2. when a sequence has matched, its subject subsequence is stored as variable $Subject; its verb as the variable $Verb and its object subsequence <DET+p> <N+p> as the variable $Object.
3. the automaton then produces the output [$Subject did not $Verb_ $Object], in which it copies the value of the variable $Subject, followed by the sequence did not, followed by the lemma of the value of the variable $Verb, followed by the value of the variable $Object.
For instance, if the text contains the sentence The lady ate these cookies, the corresponding output sequence in the negative form will be The lady did not eat these cookies. If the text contains the sentence She ate them, the corresponding output sequence will be She did not eat them. The graph in Figure 1 is overly simplistic. A more complete grammar can be seen in Figure 2 (for negation) and in Figure 3 (for auxiliary verbs).
Figure 2. Negation Transformation, main graph.
Figure 3. Graph Auxiliary.
The graph NP contains a description of various noun phrases: nouns preceded by modifiers, such as the large red pencil, genitive constructions, such as my cousin’s friends, and nouns with prepositional complements, such as the demonstration for the classroom. The noun phrase stored in $Subject is simply copied in the output of the graph, i.e., resulting negative sentence. Graph PP contains a description of prepositional phrases such as for my friends; graph VP contains a description of verbal phrases such as to eat the cookies and eating the cookies. The graph Aspect recognizes aspectual constructions such as continues to eat and stopped eating. The resulting negative structure produced by the transducer is does not continue to eat and did not stop eating respectively. The graph Auxiliary in Figure 3 recognizes verbal constructions with an auxiliary verb.
In Figure 3, the value of the special variable $THIS is the current word; <V+G> matches any verb in the Gerundive, <V+PP> matches any verb in the Past Participle and <V+INF> matches any verb in the Infinitive. The graph Neg recognizes the empty string and produces the negative adverbs never|no longer|not. The graph also recognizes sequences such as has eaten and is going to eat. The resulting negative structure produced by the transducer is has not eaten, is not going to eat, and will not eat respectively. The graph Modal recognizes verbal constructions with a modal verb may, might, must, should, and would, and produces the corresponding negative sentences. The graph Simple recognizes simple verb conjugated forms such as eat and ate and produces the corresponding negative sequences. This graph contains a special treatment for the verbs be, have, and need to produce the correct negative sequences we are not (instead of *we do not be), we have not, and its variant we do not have, and we need not with its variant we do not need. Figure 4 displays the sequences that have been automatically rewritten in the negative form by applying the grammar to the text “The Portrait of a Lady” (Henry James, 1881). Two phenomena should be described in this grammar to make it complete:
Negations using the determiner, e.g., He eats meat → He eats no meat.
Negations do not involve replacing positive aspectual verbs (continues to smoke) with a negative aspectual verb (stopped smoking). The present grammar produces only the sentence does (never|no longer|not) continue to smoke.
Figure 4. Applying the Negation Transformation.
The grammar produces only the sentence He does (never|no longer|not) eat meat. The grammar does not process verbal groups that contain an adverb, such as She is still smoking, to not produce questionable sentences, such as She is not still smoking.
For illustrations of transformational grammars in French, refer to 1. This automatic generation system has been successfully used in several applications, including automatic paraphrasing of Portuguese2,3, Arabic-French machine translation4, and generation of a semantic representation from sentences written in Italian5. However, 1 has shown several of its limitations if the goal is to implement an automatic transformational system. The description of transformations in this manner is procedural, requiring the sequential application of multiple transducers from an initial sentence to recognize or generate a complex sentence. However, this approach does not scale up, which is why procedural approaches to Computational Linguistics were abandoned in the 1970s. 6 addresses the problem of describing the behavior of chains of transformations. Sometimes, altering the order of applying a transducer within a chain of transformations can dramatically change the resulting sentence, while in other cases, it may not have a significant impact.
For example, from the original sentence He eats the cookie, the three chains of transformations [Pronoun0] [Focus0] [Preterit], [Focus0] [Pronoun0] [Preterit], and [Focus0] [Preterit] [Pronoun0] produce the same sentence: It is he who ate the cookie. However, the two chains of transformations [Negation] [Focus0] [Preterit], and [Focus0] [Negation] [Preterit] produce two different results, respectively: It is not he who ate the cookie, and It is he who did not eat the cookie. Describing classes of equivalence for chains of transformations is an unsolved problem. The sheer quantity of transducers that would need development is astronomical. Our estimate suggests that there are approximately 1,000 productive transformations in English alone. These transformations include actions such as pronominalizing an argument to alter sentences (The man ate the cookie He ate the cookie), by changing the tense (She will eat the cookie), the aspect (He goes on eating the cookie), the modality (She should eat the cookie), the focus (It is the cookie that he ate), by nominalizing the verb (She is a cookie eater), by adding intensifiers (He is a formidable cookie eater), constructing passive sentences (The cookie has been eaten), questions (Who ate the cookie?), and so on. However, the transducers shown in Figure 1 and Figure 2 cannot be applied to a negative sentence to produce the underlying elementary sentence, i.e., its semantic analysis. In a broader context, since these transformational transducers are not symmetric, one must construct 2 × 1,000 = 2,000 transducers to comprehensively describe the transformational system of the English language.
2 Declarative approach
Transducers that recognize some sentences (as their input) to produce transformed sentences (as their output) are fundamentally asymmetric, which is against the linguistic goal of describing transformations as symmetric pairs of sentences. We need a device that represents and processes both sentences of a transformation in the same way. Consider Figure 5. This transducer recognizes both sentences The boy eats meat and The boy does not eat meat. The transformational analysis of a sentence is therefore performed in two steps:
The parser applies the transducer to the sentence, e.g., The boy does not eat meat. The transducer then produces the output [Negation] and sets variable $Subject to The boy, $Verb to eat, and $Object to meat.
The generator then produces the language recognized by the transducer by exploring all its paths according to the values of the variables set at the first step. Because the variables were set during the first step of computation, their value can be computed during the generation step. For example, as variable $Verb stores the lexeme eat the symbol <V+PR> for the same variable/lexeme will be computed as eats. Any inflectional or derivation described in the system can then be applied to a given lexeme. For example, the derivation eat eater will be automatically produced by instantiating the symbol <N+Hum+s> (human noun, singular) for variable $Verb, whereas the combination of derivation and inflection eat eateries would be produced from the same variable $Verb by the symbol <N+Coll+p> (collective noun, plural).
Figure 5. Describing Transformations in a Declarative way.
Among all potentially generated sentences, one can produce the sentence that produces the [Elementary] output, i.e., The boy eats meat. The transformational analysis of the sentence is then presented as: [Negation]: The boy does not eat meat The boy eats meat. In other words, the system analyzes the sentence The boy does not eat meat as a negation of the sentence The boy eats meat. Figure 5 represents a very simplistic transformational grammar. Figure 6 presents a more complete transformational grammar that describes over 1,2 million declarative sentences transformed from the elementary sentence Joe loves Lea, including Joe has never stopped loving Lea, Joe fell madly in love with her, etc.,7.
Figure 6. A Transformational description of direct transitive sentences.
We are in the process of developing similar transformational grammars for a dozen languages. For instance, sentences produced by the grammar for the French sentence Luc aime Léa (Luc loves Lea) are shown in Figure 7.
Figure 7. Producing transformations automatically.
Many of the sentences produced by transformations, even though syntactically correct, are awkwardly phrased, e.g., “Ce n’est pas elle qui n’aura pas été l’amour fou de Luc” [It is not she who will not have been Luc’s passionate love]. They could be filtered out by some stylistic constraints, such as restraining the use of [Passive], avoiding performing more than one [Negation] in a sentence, and strategies alike. The declarative approach can also be used to automatically generate text in English from a semantic predicate represented by an RDF Turtle triplet8. The principle is to replace the description of the elementary sentence with a semantic predicate written in the RDF format: the system can then parse and generate both the predicate in its RDF form and the sentences that represent all the ways to state it in English. The same process can both generate English sentences from the RDF semantic representation and generate the RDF representation from any English sentence recognized by the grammar, i.e., compute its meaning.
3 A Case Study with eSPERTo Paraphrasing System
In this Section, we showcase how NooJ was used to create resources for the eSPERTo paraphrase generation system. The generation system allows, among other transformations, the inversion of subject-object arguments in a sentence with a reciprocal verb, such as the Portuguese verb parecer-se com (look alike, resemble), as illustrated in Figure 8.
Figure 8. eSPERTo paraphrasing system.
4.1 The Interface
When presented with a text, eSPERTo identifies word sequences that can be replaced with equivalent alternatives, known as paraphrases. In Figure 5, these identified sequences are enclosed within red square brackets. Upon hovering the cursor over them, eSPERTo highlights the sequences in green and presents a list of suggested paraphrases that can replace that sequence. In our example of the reciprocal verb in Portuguese, parecer-se com in the sentence a mulher é parecida com o irmão (the woman looks like the/her brother), eSPERTo suggests the inversion o irmão é parecido com a mulher (the brother looks like the woman) and the coordination of the two arguments a mulher e o irmão (the woman and the brother) são parecidos (the woman and the brother look alike) with the correct change of the verb to the plural form (é (is) transforms into são (are). Then, the different constructions of the sentence can also be paraphrased by replacing words with synonyms and phrases with similar phrases.
4.2 The Generative Process
To elucidate the features of a generative system like eSPERTo alongside ChatGPT, the aim is to outline the differences between their generative processes and the resulting outputs of each system. In ChatGPT (version 3.5), the user can see the generated paraphrases illustrated in Example 1 for the Portuguese sentence illustrated in Figure 4.
(1) PT - a mulher é parecida com o irmão
PARA - ?a irmã se assemelha ao irmão
PARA - a mulher possui semelhanças com o irmão
PARA - a mulher tem semelhanças com o irmão
PARA - a mulher apresenta semelhanças com o irmão
PARA - a mulher é similar ao irmão
PARA - ?a irmã tem traços semelhantes ao irmão
PARA - a mulher e o irmão possuem semelhanças
PARA - a mulher possui características que se assemelham ao irmão
PARA - a mulher e o irmão compartilham semelhanças
PARA - a mulher tem uma semelhança com o irmão
PARA - a mulher tem características parecidas com as do irmão
PARA - a mulher tem semelhanças com seu irmão
PARA - a mulher se assemelha ao irmão
PARA - a mulher possui traços semelhantes ao irmão
PARA - a mulher apresenta semelhanças com seu irmão
PARA - a mulher possui características que são parecidas com as do irmão
PARA - a mulher e o irmão compartilham semelhanças
PARA - a mulher e o irmão têm semelhanças
PARA - a mulher tem características que se assemelham ao irmão
PARA - a mulher possui semelhanças com o irmão
PARA - a mulher apresenta traços similares ao irmão
PARA - a mulher e o irmão possuem características semelhantes
PARA - a mulher possui traços que se assemelham ao irmão
The paraphrases generated from the original sentence showcase paraphrastic units that have been altered across different segments of the sentence. In eSPERTo, the linguistic processes involved in transforming the original sentence are transparent, with clearly delineated boundaries indicating where transformations can occur. This shift signifies visible paradigmatic changes. The methodology maintains consistency and has been validated by humans. Additionally, it’s worth noting that all the paraphrases produced by ChatGPT can also be generated by eSPERTo.
4.3 The resources
A significant part of the resources used to generate paraphrases is based on the lexicon-grammar theory1. More specifically, the resources include the lexicon-grammars for (i) human intransitive adjectives9, (ii) nominal predicates supported by the verb fazer (to do/to make)3, and (iii) nominal predicates supported by the verb ser de (to be of)10. Figure 9 shows an excerpt of the lexicon-grammar tables for patronymic adjectives highlighting two properties that are different for two different adjectives: açoriano (Azorean) which refers to people from the Azores, a Portuguese island, and iraquiano (Iraqi) which refers to people from the country Iraq. When a predicate (in this case, human intransitive adjective) has a certain property, it is marked ‘+’, otherwise, it is marked ‘-’ under that property. For instance, according to this lexicon-grammar, the copulative verb ser (to be) in ser açoriano cannot be replaced with the auxiliary verb tornar-se (become) nor the adjective be rewritten with a construction with the classifier noun nacionalidade (e.g., *ter nacionalidade açoriana - *to have Azorean nationality) whereas both constructions are possible with the adjective Iraqi: tornar-se iraquiano (become Iraqi) and ter nacionalidade iraquiana (to have Iraqi nationality).
Figure 9. Excerpt of the lexicon-grammar of patronymic adjectives.
The lexicon-grammars were semi-automatically converted to NooJ dictionaries by transforming each property marked with a ‘+’ into a dictionary entry attribute +property_name as highlighted in bold in Figure 10. This figure shows the converted NooJ entries of the adjectives highlighted in Figure ??6. As can be seen, property [+Vcoptornarse] and [+NclassPnacionalidade] highlighted in orange were added to iraquiano dictionary entry but not to açoriano as this predicate does not have those properties marked ‘+’ in the lexicon-grammar.
Figure 10. Converted NooJ entries from adjectives described in the lexicon-grammar highlighted in Figure 9.
In addition to converting the lexicon-grammar into NooJ dictionary entries, a few transformational properties described in the lexicon-grammars were represented in NooJ syntactic grammars following a procedural approach. These grammars generate paraphrases. In Figure 11, the grammar identifies constructions with the structure [N0 ser de DET A Npred com N1] (N0 to be of DET A Npred with N1) such as O Pedro é de uma grande parecença com a Maria (John is of a great resemblance with Mary) and generates the equivalent constructions with the arguments reversed ([N1 ser de DET A Npred com N0] – [N1 to be of DET A Npred with N0]) and coordinated ([N0 e N1 ser de DET A Npred] – [N0 and N1 to be of DET A Npred]). In the case we are illustrating, it generates the two paraphrases: A Maria é de uma grande parecença com o João (Mary is of a great resemblance with John) and O João e a Maria são de uma grande parecença [um com o outro] (John and Mary are of a great resemblance [with one another]). Moreover, it was also necessary to create a grammar to identify the coordinated arguments and then generate the constructions ([N0 ser de DET A Npred com N1] – [N0 to be of DET A Npred with N1]) and ([N1 ser de DET A Npred com N0] – [N1 to be of DET A Npred with N0]).
Figure 11. Excerpt from the grammar that generates paraphrases for the predicates ser de (to be of) that are symmetric: ARG0 and ARG1 are subgraphs that identify simple human noun phrases that are assigned to the global variables @ARG0 and @ARG1, respectively; the head of the noun phrases are also assigned respectively to global variables @N0 and @N1.
The subgraph :out_V_s creates the corresponding paraphrases when @N1 is singular, and :out_V_p creates the paraphrases when @N1 is plural and for the coordinated arguments so that the subjects @ARG1 or @ARG0 e @ARG1 (@ARG0 and @ARG1)agree in number with the support verb (@V).
However, as pointed out in 2, which includes a detailed description of how the lexicon-grammar information was integrated into NooJ, eSPERTo will benefit greatly by using the declarative approach as for any set of n sentences that are paraphrases of each other one needs to develop only one graph stating that the n paraphrases are equivalent whereas with the procedural approach, it is necessary to create n graphs with (n -1) paths in a total of n x (n -1) paths, because each graph identifies one sentence only in the set of paraphrases and generates all other n-1 paraphrases in the set.
In addition to the lexicon-grammar resources already in existence, INESC-ID’s scholars introduced a new concept and coined it "paraphrasary"11,12, which encompasses advanced expansions of dictionaries incorporating resources spanning multiword and phrasal levels. The Multi3Generation COST Action has propelled research in monolingual and multilingual applications, particularly through the development of paraphrasaries. These resources aim to address the gap in linguistically sophisticated tools. Experimental tests, aided by alignment tools, have informed the creation of guidelines for aligning paraphrases and translations to improve semantic equivalence. Expanding the volume of paraphrastic units and extending alignment processes to more creative text types are critical endeavors. Collaborative efforts in building paraphrasaries stand as a significant advancement in linguistic research, promising enhancements in NLP applications.
4 Conclusion
The linguistic objective of describing natural languages is a valuable scientific pursuit, regardless of any immediate or short-term applications. Accessing an exhaustive and precise description of a language benefits all NLP software applications. Simply enumerating sentences within a language is insufficient; it is crucial to establish connections between sentences derived through transformations, such as active and passive versions or affirmative and negative forms, among others. This article presents two methods for formalizing these connections: the procedural approach and the declarative approach. The procedural method uses simple transducers to generate transformed sentences from elementary ones, taking an elementary sentence as input and producing the corresponding transformed sentence as output. The procedural approach, exemplified by eSPERTo, has demonstrated success in automatically producing paraphrases (using monolingual paraphrasaries) and translations (when associated with bilingual dictionaries). The declarative approach employs a single "neutral" grammar to represent both elementary and transformed sentences. Future iterations of eSPERTo could explore adopting the declarative approach to represent equivalent transformations within a single grammar. "Neutral" grammars can function in both "parsing" and "generating" modes using a single computational device to process (parse or generate) texts. Their advantage lies in enabling linguists to describe transformations in a purely declarative manner. A comprehensive transformational grammar of a language would be the linguistic equivalent of the human genome project. The NooJ community has made strides in developing transformational grammars for numerous languages over several years. These grammars can further improve the generative capabilities of paraphrase generation in eSPERTo and other NLP applications, regardless of their model type.
Data and software availability
The procedural approach is a methodology implemented in NooJ, a free and open-source software that can perform the functions described in the article (https://nooj.univ-fcomte.fr).  has a large international multilingual community that uses and shares results at the annual NooJ internal conference. To ensure the reproducibility of our research, we provide detailed descriptions of the materials and methods used, allowing others to replicate our work. All data supporting the findings of this study are available within the article.  relevant materials are already accessible through the website and other channels.  tutorial and resources on the NooJ website, known for creating high-quality linguistic resources for many languages. There is no need for an additional description of extended data or novel analysis code.  Tutorial is included. Sufficient and relevant references to the most pertinent publications for this article are provided  and other visual data were created using software developed by the authors. Any additional inquiries regarding data access can be directed to the corresponding author.  has a large international multilingual community that uses and shares results at the annual NooJ internal conference. To ensure the reproducibility of our research, we provide detailed descriptions of the materials and methods used, allowing others to replicate our work. All data supporting the findings of this study are available within the article. relevant materials are already accessible through the website and other channels.  tutorial and resources on the NooJ website, known for creating high-quality linguistic resources for many languages. There is no need for an additional description of extended data or novel analysis code.  Tutorial is included. Sufficient and relevant references to the most pertinent publications for this article are provided  and other visual data were created using software developed by the authors. Any additional inquiries regarding data access can be directed to the corresponding author.
Nooj Tutorials and resources can be accessed as by the following link:
https://nooj.univ-fcomte.fr/tutorials.html
eSPERTo paraphrase generation system -
eSPERTo code and resources can be accessed freely for research purposes by requesting access to the principal investigator of the eSPERTo project, Anabela Barreiro (anabela.barreiro@inesc-id.pt)
Author contributions
Max Silberztein, Multi3Generation’s Grant Coordinator, wrote most of Section 1, and Section 2 and Section 3. Cristina Mota, Multi3Generation member, wrote part of Section 4.
Anabela Barreiro, Multi3Generation’s chair, coordinated the article, organised the structure of the manuscript, wrote the Abstract, Plain Language Summary, Introduction, part of Section 4, Subsection 4.2, part of Subsection 4.3, Conclusion, and revised the entire article.
Acknowledgment
This article is based upon work from COST Action Multi3Generation (CA18231) supported by COST (European Cooperation in Science and Technology).
Faculty Opinions recommendedReferences
- 1.
Gross M:
Les bases empiriques de la notion de prédicat sémantique.
Langages.
Included in a thematic issue: Formes syntaxiques et prédicats sémantiques. 1981; 15(63): 7–52. Reference Source
- 2.
Barreiro A, Mota C, Baptista J, et al.:
Linguistic resources for paraphrase generation in portuguese: a lexicon-grammar approach.
Lang Resour Eval.
2022; 56(1): 1–35. Publisher Full Text
- 3.
Chacoto L:
O verbo fazer em construções nominais predicativas. PhD thesis, Universidade do Algarve, 2005. Reference Source
- 4.
Essid M, Fehri H:
A semantico-syntactic disambiguation system of Arabic movement and speech verbs and their automatic translation to French using NooJ. 12th International Conference, NooJ 2018, Palermo, Italy, June 20– 22, 2018. Revised Selected Papers. 2019; 987. : 167–179. Publisher Full Text
- 5.
Vietri S:
The annotation of the predicate-argument structure of transfer nouns. Cambridge Scholars Publishing, 2013; 88–101. Reference Source
- 6.
Silberztein M:
Formalizing natural languages: the NooJ approach. 2016. Reference Source
- 7.
Silberztein M:
Joe Loves Lea: transformational analysis of direct transitive sentences.
Automatic Processing of Natural-Language Electronic Texts with NooJ.
2016; 607: 55–65. Publisher Full Text
- 8.
Silberztein M:
From FOAF to english: linguistic contribution to web semantics. 2017; 1–9. Reference Source
- 9.
Carvalho P:
Análise e representacão de construcões adjectivais para processamento automático de texto. Adjectivos intransitivos humanos. 2007. Reference Source
- 10.
Baptista J:
Sintaxe dos nomes predicativos com verbo-suporte SER DE. Fundação para a Ciência e a Tecnologia/Fundação Calouste Gulbenkian, 2005. Reference Source
- 11.
Barreiro A, Rebelo-Arnold I, Mota C:
Parafrasário: a variety-based paraphrasary for portuguese. In: Proceedings of the 17th NooJ International Conference 2023 – Book of Abstracts. Zadar, Croacia, University of Zadar, Department of Classical Philology and Department of Information Sciences, 2023; 44–45. Reference Source
- 12.
Barreiro A, Mota C:
A multilingual paraphrasary of multiwords. In: Proceedings of the 1st International Workshop on Multilingual, Multimodal and Multitask Language Generation (Multi3Generation).
Tampere, Finland, The Association for Computational Linguistics. 2023; 47–56. Reference Source
