This paper presents the initial stages of a project focused on coreference and anaphora resolution in Latin texts. By building a corpus enhanced with coreference/anaphora annotation, the project wants to explore empirically a layer of metalinguistic analysis that has not been yet extensively investigated in linguistic resources and natural language processing for Latin. After reviewing the related work on this NLP task, the paper discusses annotation criteria and data analysis, providing examples about a few issues that emerged during the annotation process.
also confirmed by those selected for the CoNLL shared task on modeling unrestricted coreference in OntoNotes Coreference (henceforth CR) and anaphora (henceforth [9, 10], as well as by the NXT-format Switchboard CorAR) resolution are often treated as a single, yet diverse, pus [11]. In addition, some treebanks feature CR/AR, task in NLP. To understand the diference between CR encompassing a wide range of languages, including Enand AR, it is necessary to distinguish between the con- glish and Czech [12], German [13], Japanese [14], Italian cept of “mention” and that of “entity”. A mention is [15], Spanish and Catalan [16]. To the best of our knowldefined as an instance of reference to an object, while edge, there is no specific Latin corpus enriched with an entity is the object to which a mention refers in a CR/AR. The only currently available texts that include text. CR consists in finding in a text all mentions of this layer of annotation come from Latin treebanks. The (strictly speaking, real-world) entities such as persons or FIR-2013 project mentioned above built a CR-annotated organisations, regardless of their textual representation. dataset including works by Sallust, Caesar and Cicero Instead, in AR the interpretation of a mention (known as (taken from the Latin Dependency Treebank [17]), and “anaphora” or “cataphora”, e.g., a pronoun) depends on by Thomas Aquinas (from the Index Thomisticus Treeanother mention present in the text, whether antecedent bank [18]). However, the selection of texts in this dataset or following in the word order. If both mentions refer to is quite unbalanced as for both literary genres and authe same entity, they are considered to be coreferential, thors. Out of the more than 45,000 total annotated tokens, which makes AR and CR closely bound to each other. about 27,000 are taken from Thomas Aquinas’ Summa Since ana-/cataforic relations are present in the text, the contra Gentiles, and more than 10,000 are from Sallust’s need of world knowledge in AR is minimal. In contrast, In Catilinam. This given, our project wants to create a CR has a much broader scope: co-referential terms can more balanced dataset by increasing and diferentiating have completely diferent grammatical properties and/or the quantity of annotated texts for both Classical and functions (e.g., diferent gender and part of speech) and Late Latin. yet, by definition, they can refer to the same entity.
In NLP, the CR task is usually not meant in a strict sense, as it consists in finding all mentions of each entity 3. Building CorefLat in a text regardless of their relation to the real world.
Accordingly, our project adopts this same interpretation 3.1. Annotation Criteria and Data of the CR task [4]. Selection
Since the 1960s, coreference and anaphora resolution has been a central topic in NLP studies, but it was con- To create a resource that adheres to the most unified and sidered a dificult task, typically requiring the use of widely shared annotation criteria for CR/AR, the annosophisticated knowledge sources and inference proce- tation style of CorefLat resembles the one developed for dures. In 1983, Roberto Busa pointed out the absence of the GUM corpus and follows the recommendations proresources and tools for pronoun coreference resolution: posed by the (ongoing) Universal Anaphora (UA) project4, “[...] avete mai incontrato tavole e concordanze comput- which aims to create, gather, and distribute harmonized erizzate nelle quali il programma automaticamente abbia resources for CR/AR. [...] collegato i pronomi alle forme di cui sono vicari?” [5, While building CorefLat, we decided to focus on a 7.2]3. subset of the diferent types of coreference and ana
Like for other NLP tasks, during the 1990s research on /cataphora prescribed by the GUM and UA recommendaCR/AR gradually shifted from heuristic approaches to tions. The types that we selected are listed below: machine learning approaches, thanks to the public availability of annotated corpora produced for the aims of shared tasks dedicated to coreference resolution, such as Message Understanding Conference (MUC) conferences [7], and Automatic Content Evaluation (ACE) Program conferences [8]. These corpora mainly include news article and newswire texts in English. The ACE corpus also features Arabic and Chinese texts from web-blogs and telephone conversations. The tendency to focus coreference and anaphora annotation on newspaper texts is • anaphoric pronouns referring back to something: domine qui et semper vivis (Aug. Conf. 1.6.8) ‘Lord (you) who live for ever’; • cataphoric pronouns referring forward to something: invocat te, domine (Aug. Conf. 1.1.1) ‘invokes you, Lord’; • content-rich lexical item - coreferring the same lexical mention: laudes tuae, domine, laudes tuae per scripturas tuas suspenderent palmitem cordis mei (Aug. Conf. 1.17.27) ‘Your praises, Lord, your praises throughout your Scriptures would have supported the vine shoot of my heart’; 3“[...] have you ever come across computerized tables and concordances in which the programme automatically [...] connects pronouns with the nouns that they represent?”. Translation taken from [6, 137-138].
4https://universalanaphora.github.io/UniversalAnaphora/ • split antecedents - the referred items are more content-rich entity is concerned in this coreference than one: an vero caelum et terra, quae fecisti relation. Moreover, it should be noted that sometimes et in quibus me fecisti, capiunt te? (Aug. Conf. the entity is not explicitly expressed in the text. To 1.2.2) ‘heaven and earth, which you made, and address this issue, we create external entities to which in which you made me, encompass you?’. the respective mentions are linked by tagging. For instance, in example (3), the pronoun nos ‘we’ refers to the two lovers in Plautus’ comedy Curculio, namely the girl Planesium and the boy Phaedromus, whose names are not explicitly mentioned in the sentence for economy’s sake, as the two characters are present on stage and pronounce these lines themselves.
to address the fundamental aim of the two-year long funded project, namely building and distributing a Latin corpus enhanced with coreferential annotation, which is not yet available for this language.
Texts are annotated manually by two independent annotators, using the Content Annotation Tool (CAT)[19], (3) quo usque, quaeso, ad hunc modum / inter nos formerly known as the CELCT Annotation Tool, which amore utemur semper surrupticio? (Pl. Curc. 1, 204-205) was created specifically for textual coreference annota- ‘How much longer, please, will we always conduct our tion. The tool is highly customizable, making it possible, love afair in secret?’ for instance, to distinguish between annotations of mentions and those of entities. (Meta)data are saved in In such a case, we tag the mention nos as linked XML and are then converted in CoNLL-U Plus following to the entities “Planesium” and “Phaedromus” that are the recommendations of the UA initiative5. created external to the text.
In CorefLat, coreferences are not annotated as chains, The annotation task is performed on a collection of but rather as relations. In a coreference relation two Latin texts already enriched with lemmatization and Partelements are involved: the one referring (mention) of-Speech (PoS) tagging and linked to the LiLa Knowland the one referred (entity). In our annotation, each edge Base. The following texts were chosen according to mention points directly to the one entity it refers to, selection criteria aimed to ensure a suficiently represenrather than to any previous mention of the same entity. tative and balanced corpus as for both literary genre and Consider the example in (1). era. (1) Magnus es, Domine, et laudabilis valde. Magna virtus tua et sapientiae tuae non est numerus. (Aug. Conf. 1.1.1) ‘Great are you, O Lord, and surpassingly worthy of praise. Great is your goodness, and your wisdom is incalculable’6.
In sentence (1), we identify two coreference relations: the first one involves the mention tua and the entity Domine, and the second one involves the mention tuae and the same entity Domine. Typically, the referred element is a noun, nevertheless it happens to get through cases where the referred entity is represented by a function word, such a pronoun, like in example (2): (2) nec valerem quae volebam omnia nec quibus volebam omnibus. (Aug. Conf. 1.8.13) ‘I was incapable of achieving all that I wanted, and by all that I wanted.’ In (2), the relative pronoun quae refers to the quantifying pronoun omnia, like quibus refers to omnibus in the reminder of the sentence. Since omnis ‘all’ (lemma of both omnia and omnibus) is a function word, no
main/documents/UA_CONLL_U_Plus_proposal_v1.0.md 6English translations of Latin examples are taken, with minor changes, from [20] (Augustine) and [21] (Plautus). • Classical Latin: data are excerpted from the Opera Latina corpus by LASLA7, an extensive collection of approximately 1.7 million words from over 130 lemmatized and morphologically tagged Classical and Late Latin texts8. • Late Latin: data are taken from the text of Augustine’s Confessiones provided by The Latin Library9.
3.2. Results So far, we annotated the following excerpts: the first book from Augustine’s Confessiones, a philosophical prose text, and a comedy of Plautus: Curculio. The workload was split equally between the two annotators; however, the last 50 sentences of the first book of Augustine’s Confessiones were annotated by both annotators to measure 7https://lasladb.uliege.be/OperaLatina/ 8The Opera Latina corpus in the LiLa Knowledge Base is available at https://lila-erc.eu/data/corpora/Lasla/id/corpus. 9http://www.thelatinlibrary.com. The text is available in LiLa at https://lila-erc.eu/lodview/data/corpora/CIRCSELatinLibrary/id/ corpus/Confessiones their agreement. Inter-annotator agreement was cal- 3.3. Annotation Issues culated through the Dice coeficient similarity metric, which is widely adopted in NLP [22, 23]. Its value ranges from 0 to 1, with 1 indicating that two sets are identical and 0 meaning that they have no overlap. Once evaluated that the annotated markables span the same tokens for the two annotators in all cases, we calculated the similarity values as for entities (0.817) and mentions (0.824), which are comparatively highly acceptable for this task [24, 25, 26]. Additionally, the Cohen’s Kappa coeficient was measured, yielding the following agreement values for each markable class: for the markable class ‘mention’ the resulting value is 0.8139902, whereas for the markable class ‘entity’, the value obtained is 0.8118851.
Table 1 presents the data derived from the analysis of the two texts. To highlight the quantitative significance of the coreference phenomenon, it shows the total number of tokens in the texts analyzed, along with the number of tokens involved in coreference relations. Additionally, the table shows the total number of coreference relations, and their respective entities and mentions. The
of annotation issues. On one hand, we address a problematic case regarding the application of our annotation scheme on the data, which was the primary reason for disagreement between the two annotators (example 4). On the other hand, we present two cases that highlight the fundamental role of context (example 5) and of the literary genre (example 6) for the coreference resolution task. The limited number of cases presented below is consistent with our prior decision to restrict the scope of annotation to only a subset of coreferential phenomena. We hypothesize that expanding the range of annotated coreference types or enlarging the corpus of annotated texts (in terms of quantity and literary genre) would lead to greater annotation challenges.
Starting from the first annotation issue, the most relevant disagreement between the two annotators concerns how to link mentions that are distant in the text from the entity they refer to. Example (4) shows a representative case of this type of disagreement. tokens involved in a coreference relation account for the 12.16 percent of the total in Confessiones, while in Curculio they represent the 16.7 percent of the total. In both cases the percentages exceed the data produced by the FIR project, where the phenomenon concerns approximately the 8 percent of the tokens of the Latin texts annotated therein. The table clearly indicates that Curculio exhibits a greater number of coreferences despite having a lower total number of tokens. This diference is statistically significant: the chi-squared test performed on these data yielded a chi-squared statistic of 49.18 and a p-value lower than 0.00001. Given that the p-value is lower than the conventional alpha level of 0.05, coreference relations vary significantly from a statistic point of view in Confessiones and in Curculio. The coreference phenomenon is indeed widespread in the language of Plautus’s theatre. This may be due to the fact that Plautus’s language mimics, to some extent, everyday spoken language. Furthermore, the presence of numerous dialogues, where speakers often interrupt each other’s turns, implies frequent references to the recipients with whom the characters interact. The text structure, characterized by numerous allocutions, also contributes to the high number of coreferences. (4) Bonus ergo est qui fecit me, et ipse est bonum meum, et illi exulto bonis omnibus quibus etiam puer eram. Hoc enim peccabam, quod non in ipso sed in creaturis eius me atque ceteris voluptates, sublimitates, veritates quaerebam, atque ita inruebam in dolores, confusiones, errores. (Aug. Conf. 1.20.31) ‘Therefore the one who made me is good, and he himself is my good, and I rejoice in him for all the good things of which I consisted even in childhood. This was my sin: I sought pleasures, exaltations, truths not in he himself but in his creations, which is to say, in myself and other things’.
God, which is explicitly expressed six sentences above in the text. The reader has no dificulty decoding these pronouns because the first-person narrator is discussing his relationship with God, to whom he is constantly referring. Therefore, it is not necessary to explicitly state the entity in every sentence.
The sentence in (4) can be annotated in two distinct ways: each pronoun can either be directly linked to the entity ‘God’ within the text, or be linked to the first pronoun concerned in (4) (qui), which gets then linked to the external entity ‘God’. During the annotation process, the two annotators diverged: one selected the former method, while the other opted for the latter. There is no upper limit to the number of sentences after which a mention cannot be associated with the entity to which it refers [27]. When CR and AR first emerged as NLP tasks, there were concerns that machines could not yield acceptable results if the mention and the entity were too distant from each other [28]. However, contemporary meth- refer to the same entity. This case clearly demonstrates ods achieve satisfactory results even with long-distance the importance of understanding both the context and coreference, exceeding 200 sentences [29]. Additionally, the specific narrative techniques of the textual genre in given that we focus on literary texts, which feature long- order to efectively resolve coreferences. distance coreferences more frequently than other textual types [30], it is imperative that we devote particular attention to this specific type of coreference. The two options 4. Conclusion and Future Work chosen by the annotators are both equally valid. To harmonize the annotation process, we decided to link the mention to the external entity beyond a certain threshold, which was set at five sentences 10.
Sentence (5) from Plautus’ Curculio exemplifies another challenging case of ambiguity, which further complicates the annotation process: (5) Pal.: Quid? tu te pones Veneri ieientaculo? Phaed.: Me, te atque hosce omnis. (Pl. Curc. 1, 73-74) Pal.: ‘What? You’ll ofer yourself a breakfast to Venus?’ Phaed.: ‘Yes, myself, yourself, and all these here.’ As is typical in theatrical texts, much is left to the audience’ inference. In this instance, the actor’s gestures serve to disambiguate the phrase hosce omnis, which could refer either to the group of slaves accompanying the character Phaedromus or to the audience itself [31, 32, 33]. The annotators decided to follow the interpretation provided by Paratore [34], according to whom, hosce omnis refers to the audience. In this example, an agreement in gender and number between the mentions and the potential antecedents inferred from the context can be observed. Disambiguating the antecedent not only requires understanding the text but also knowing the specific characteristics of the literary genre concerned.
Another case in which the importance of literary genre and knowledge of context becomes evident is as follows. (6) Cvrc.: [...] Lyconem quaero tarpezitam. Lyc.: Dic mihi, quid eum nunc quaeris? (Pl. Curc. 3, 406- 407): Cvurc.: ‘I’m looking for the banker Lyco.’ Lyc.: ‘Tell me, why are you looking for him now?’
ters, Curculio and Lyco, plays on a comedic ambiguity: Curculio knows he is speaking to Lyco, while Lyco believes that Curculio is unaware of his identity. When Curculio asks to speak with Lyco, Lyco responds by speaking about himself in the third person, thereby concealing his identity. For this reason, both the first-person pronoun ‘mihi’ and the third-person pronoun ‘eum’
In this paper, we provide an overview of the current state of a project aimed to build a Latin corpus enhanced with coreference and anaphora resolution. We detailed the annotation criteria and discussed a few annotation challenges, highlighting how this annotation layer necessitates a profound interaction among various fields of expertise, including linguistics, textual criticism, and literature.
In the near future, our aim is to expand the annotated corpus and to further extend the evaluation of interannotator agreement by incorporating the metrics as those proposed by Kopeć and Ogrodniczuk [35], such as the MUC score [36]. Once a suficiently large dataset will be available, NLP will be concerned too, as we plan to exploit the annotated dataset to train and evaluate a stochastic model in supervised fashion to perform automatic CR/AR of Latin, usable also in NLP pipelines like, for instance, UDPipe [37] and Stanza [38]. We expect such a model to prove helpful to provide the Latin treebanks currently available in the Universal Dependencies (UD) initiative [39] with a layer of so-called enhanced dependencies, which also includes coreference and anaphora resolution. This would position Latin on an equal footing with other contemporary languages for which CR/AR annotations are also publicly accessible in treebanks [40] 11. Given that one of the UD Latin treebanks, the Index Thomisticus Treebank, is already published as Linked Data in the LiLa Knowledge Base [41], having the treebank enriched with enhanced dependencies will require to model and publish therein the metadata about CR/AR.
The contribution of our project can also be considered within the broader context of NLP task on Latin. For instance, the corpus enriched with coreference annotations could enhance a task such as Emotion Polarity Detection, which was one of the shared tasks at the last edition of the evaluation campaign EvaLatin 2024. In the long term, a follow-up of the project will consist in building further textual datasets that feature other layers of coreferential annotation recognized by the GUM framework, such as appositive, attributive, and predicative coreferences, along with discourse deixis, and non-proper coreferences.
Finally, given the current spread of Large Language Models and their highly promising accuracy rates on a wide already models for Latin, such as the Latin BERT [42]. tilingual unrestricted coreference in ontonotes, in: Joint conference on EMNLP and CoNLL-shared task, 2012, pp. 1–40. 5. Acknowledgements [11] S. Calhoun, J. Carletta, J. M. Brenier, N. Mayo, D. Jurafsky, M. Steedman, D. Beaver, The nxt-format This contribution is funded by the PRIN 2022 project "Tex- switchboard corpus: a rich resource for investigattual Data and Tools for Coreference Resolution in Latin" ing the syntax, semantics, pragmatics and prosody (CUP J53D23013680008), a project carried out jointly by of dialogue, Language resources and evaluation 44 the Università Cattolica del Sacro Cuore in Milan and by (2010) 387–419. the University of Udine. [12] A. Nedoluzhko, M. Novák, S. Cinková, M. Mikulová,
J. Mírovský, Coreference in Prague Czech-English References Dependency Treebank, in: N. Calzolari, K. Choukri,
T. Declerck, S. Goggi, M. Grobelnik, B. Maegaard, [1] M. Passarotti, F. Mambrini, G. Franzini, F. M. Cec- J. Mariani, H. Mazo, A. Moreno, J. Odijk, S. Piperidis chini, E. Litta, G. Moretti, P. Rufolo, R. Sprugnoli, (Eds.), Proceedings of the Tenth International ConInterlinking through lemmas. the lexical collection ference on Language Resources and Evaluation of the lila knowledge base of linguistic resources for (LREC’16), European Language Resources Associalatin, Studi e Saggi Linguistici 58 (2020) 177–212. tion (ELRA), Portorož, Slovenia, 2016, pp. 169–176. [2] M. Passarotti, From syntax to semantics. first steps URL: https://aclanthology.org/L16-1026.
towards tectogrammatical annotation of latin, in: [13] E. Hinrichs, S. Kübler, K. Naumann, H. Telljohann, Proceedings of the 8th Workshop on Language J. Trushkina, et al., Recent developments in linTechnology for Cultural Heritage, Social Sciences, guistic annotations of the TüBa-D/Z treebank, Uniand humanities (LaTeCH), 2014, pp. 100–109. versitätsbibliothek Johann Christian Senckenberg, [3] T. Berners-Lee, J. Hendler, O. Lassila, The semantic 2004.
web, Scientific american 284 (2001) 34–43. [14] R. Iida, M. Komachi, K. Inui, Y. Matsumoto, Annotat[4] R. Sukthanker, S. Poria, E. Cambria, ing a japanese text corpus with predicate-argument R. Thirunavukarasu, Anaphora and corefer- and coreference relations, in: Proceedings of the ence resolution: A review, Information Fusion 59 linguistic annotation workshop, 2007, pp. 132–139. (2020) 139–162. [15] A. Minutolo, R. Guarasci, E. Damiano, G. De Pietro, [5] R. Busa, Trent’anni d’informatica su testi: a che H. Fujita, M. Esposito, A multi-level methodology punto siamo? quali spazi aperti alla ricerca?, in: for the automated translation of a coreference resConvegno su L’Università e l’evoluzione delle Tec- olution dataset: an application to the italian lannologie Informatiche, volume 1, CILEA, Milano, guage, Neural Computing and Applications 34 Italy, 1983, pp. 7.1–7.4. (2022) 22493–22518. [6] J. Nyhan, M. Passarotti, One Origin of Digital Hu- [16] M. Recasens, M. A. Martí, Ancora-co: Coreferenmanities: Fr Roberto Busa in His Own Words, tially annotated corpora for spanish and catalan, Springer, 2019. Language resources and evaluation 44 (2010) 315– [7] N. A. Chinchor, Overview of MUC-7, in: Sev- 345.
enth Message Understanding Conference (MUC- [17] D. Bamman, G. Crane, The design and use of a latin 7): Proceedings of a Conference Held in Fairfax, dependency treebank, in: Proceedings of the Fifth Virginia, April 29 - May 1, 1998, 1998. URL: https: Workshop on Treebanks and Linguistic Theories //aclanthology.org/M98-1001. (TLT2006), Citeseer, 2006, pp. 67–78. [8] G. R. Doddington, A. Mitchell, M. A. Przybocki, [18] M. Passarotti, The project of the index thomisticus L. A. Ramshaw, S. M. Strassel, R. M. Weischedel, treebank, in: Digital Classical Philology, De Gruyter The automatic content extraction (ace) program- Saur, 2019, pp. 299–320. tasks, data, and evaluation., in: Lrec, volume 2, [19] V. B. Lenzi, G. Moretti, R. Sprugnoli, Cat: the celct Lisbon, 2004, pp. 837–840. annotation tool., in: LREC, 2012, pp. 333–338. [9] S. Pradhan, L. Ramshaw, M. Marcus, M. Palmer, [20] W. W. Augustine, Confessions, Vol. 2: Books 9-13 R. Weischedel, N. Xue, Conll-2011 shared task: (Loeb Classical Library, No. 27), 1912.
Modeling unrestricted coreference in ontonotes, in: [21] P. Nixon, et al., Plautus, Vol. II: Casina. The Casket Proceedings of the fifteenth conference on com- Comedy. Curculio. Epidicus. The Two Menaechputational natural language learning: shared task, muses (Loeb Classical Library), William Heine2011, pp. 1–27. mann; GP Putnam’s Sons, 1917. [10] S. Pradhan, A. Moschitti, N. Xue, O. Uryupina, [22] L. R. Dice, Measures of the amount of ecologic Y. Zhang, Conll-2012 shared task: Modeling mul- association between species, Ecology 26 (1945) 297– 302. Association for Computational Linguistics [23] T. Sorensen, A method of establishing groups of 11 (2023) 922–940. URL: https://doi.org/10. equal amplitude in plant sociology based on similar- 1162/tacl_a_00581. doi:10.1162/tacl_a_00581. ity of species content and its application to analyses arXiv:https://direct.mit.edu/tacl/article-pdf/doi/10.11 of the vegetation on danish commons, Biologiske [37] M. Straka, J. Hajic, J. Straková, Udpipe: trainable skrifter 5 (1948) 1–34. pipeline for processing conll-u files performing tok[24] K. B. Cohen, A. Lanfranchi, M. J.-y. Choi, M. Bada, enization, morphological analysis, pos tagging and W. A. Baumgartner, N. Panteleyeva, K. Verspoor, parsing, in: Proceedings of the Tenth International M. Palmer, L. E. Hunter, Coreference annotation Conference on Language Resources and Evaluation and resolution in the colorado richly annotated full (LREC’16), 2016, pp. 4290–4297. text (craft) corpus of biomedical journal articles, [38] S. N. Group, et al., Stanza–a python nlp package for BMC bioinformatics 18 (2017) 1–14. many human languages, 2018. [25] I. Hendrickx, G. Bouma, F. Coppens, W. Daelemans, [39] M.-C. De Marnefe, C. D. Manning, J. Nivre, D. ZeV. Hoste, G. Kloosterman, A.-M. Mineur, J. Van man, Universal dependencies, Computational linDer Vloet, J.-L. Verschelde, A coreference corpus guistics 47 (2021) 255–308.
and resolution system for dutch., in: LREC, 2008. [40] V. Ng, Supervised noun phrase coreference re[26] A. Nedoluzhko, J. Mírovsky`, P. Pajas, The coding search: The first fifteen years, in: Proceedings scheme for annotating extended nominal corefer- of the 48th annual meeting of the association for ence and bridging anaphora in the prague depen- computational linguistics, 2010, pp. 1396–1411. dency treebank, in: Proceedings of the Third Lin- [41] F. Mambrini, M. Passarotti, G. Moretti, M. Pelleguistic Annotation Workshop (LAW III), 2009, pp. grini, The index thomisticus treebank as linked 108–111. data in the lila knowledge base, in: Proceedings of [27] R. Simone, Fondamenti di linguistica, volume 9, Lat- the Thirteenth Language Resources and Evaluation erza Bari, 1990. Conference, 2022, pp. 4022–4029. [28] T. McEnery, I. Tanaka, S. Botley, Corpus annotation [42] D. Bamman, P. J. Burns, Latin bert: A contextual lanand reference resolution, Operational Factors in guage model for classical philology, arXiv preprint Practical, Robust Anaphora Resolution for Unre- arXiv:2009.10053 (2020).
stricted Texts (1997). [29] H.-L. Trieu, A.-K. D. Nguyen, N. Nguyen, M. Miwa,
H. Takamura, S. Ananiadou, Coreference resolution in full text articles with bert and syntax-based mention filtering, in: Proceedings of the 5th workshop on BioNLP open shared tasks, 2019, pp. 196–205. [30] R. Thirukovalluru, N. Monath, K. Shridhar, M. Zaheer, M. Sachan, A. McCallum, Scaling within document coreference to long texts, Findings of the Association for Computational Linguistics: ACL
IJCNLP 2021 (2021) 3921–3931. [31] L. Cappiello, Un commento al curculio di plauto
(vv. 1-370) (2015). [32] G. Monaco, T. M. Plauto, Teatro di Plauto: Il Cur
culio. I, Istituto editoriale cultura europea, 1963. [33] T. H. Gellar-Goad, Plautus’ curculio and the case of the pious pimp, Roman Drama and its Contexts 34 (2016) 231. [34] T. M. Plautus, E. Paratore, Il gorgoglione:(Il Gorgoglione). Testo latino con traduzione a fronte, Sansoni, 1958. [35] M. Kopeć, M. Ogrodniczuk, Inter-annotator agreement in coreference annotation of polish, Advanced Approaches to Intelligent Information and Database
Systems (2014) 149–158. [36] B. Zheng, P. Xia, M. Yarmohammadi, B. V.
Durme, Multilingual Coreference Resolution in Multiparty Dialogue, Transactions of the