3.1Units of analysis

For this study we have selected named entities as our units of analysis. NEs are special types of terms that identify real-world “objects”, understood in a broad sense to encompass persons (Ursula von der Leyen), geopolitical locations, such as cities, countries, and states (Brussels, Belgium, UK), non-geopolitical locations, such as mountain ranges, seas, rivers, etc. (Danube), organizations and governing bodies (the International Court of Justice, the European Agency for Medicine EMA), buildings and other infrastructure (Westminster Abbey), and products (AstraZeneca), as well as dates (13 April, 2021), times (hour), numbers (third, 40.5), quantities and percentages (kilowatt, 3%), currency (Sterling pound, £) and even languages (English), among others.

Named entities, especially those referring to persons, locations and organizations, behave exactly like terms11.In fact, named entities have also been referred to as “system-bound” or “culture-bound” terms in the legal translation literature (Vigier-Moreno and Sánchez Ramos 2017Vigier-Moreno, Francisco Javier and María del Mar Sánchez Ramos 2017 “Using parallel corpora to study the translation of legal system-bound terms: The case of names of English and Spanish courts.” In Computational and Corpus-Based Phraseology. Second International Conference, Europhras 2017 London, UK, November 13–14, 2017 Proceedings, edited by Ruslan Mitkov, 260–273. Cham: Springer. ). However, despite drawing attention to common problems which professionals could face when translating these entities, the literature does not explore the possible relations between single or multi-word entities, collocations, and formulaic structures in the legal and institutional discourse. and present similar challenges for their systematic study: homographic pairs, variant spellings, morphological inflection, phraseological patterns, and the possibility of being referred to as multi-word expressions or acronyms (Jacquet et al. 2019Jacquet, Guillaume, Maud Ehrmann, Jakub Piskorski, Hristo Tanev, and Ralf Steinberger 2019 “Cross-lingual linking of multi-word entities and language-dependent learning of multi-word entity patterns.” In Representation and Parsing of Multiword Expressions: Current trends, edited by Yannick Parmentier and Jakub Waszczuk, 269–297. Berlin: Language Science Press. ).

For the formulaicity analysis, we will use the classification proposed by Biel (2014Biel, Łucja 2014 “Phraseology in legal translation: A corpus-based analysis of textual mapping in EU law.” In The Ashgate Handbook of Legal Translation, edited by Le Cheng, King Kui Sin & Anne Wagner, 177–192. , 178–181), which depicts a phraseological continuum with fuzzy boundaries between five categories, ranging from the global textual level to the local microlevel. Examples (in italics) illustrate various phraseological patterns (in bold) with NEs (underlined).

• Text-organizing patterns are repetitive global textual sequences which are often prescribed in drafting guidelines. They form a matrix of a legal text, emphasizing its ritualized nature. Typical text-organizing patterns include the title of the document, citations, transitions between sections, enacting formulas, amending formulas, and closing formulas, e.g.: The meeting opened at 10.06 on Wednesday, 19 February 2020, with Ms Yana Toom, (2nd Vice – Chair) presiding .

• Grammatical patterns are genre-specific, recurrent, and express deontic modality, if-then mental models of legal reasoning and other conditional clauses, purpose clauses, the passive voice, and other impersonal structures. E.g.: PETI should not draft an opinion to the AFET, El artículo 29 de la CDPD ‎debe entenderse junto con el artículo 9 (Accesibilidad) [Article 29 of the CRPD should be read in conjunction with Article 9 (Accessibility)].

• Term-forming patterns are collocates of a generic term which form more specific multi-word terms of varying degrees of terminologicality. The typical, most productive term-forming patterns tend to be Adj + N and N + N, but in practice multi-word terms may be structurally very complex, e.g.: European Commission , cuenca del Mar Menor [(The) Mar Menor basin].

• Term-embedding collocations22.The term collocation used in this paper follows Biel’s approach (2014Biel, Łucja 2014 “Phraseology in legal translation: A corpus-based analysis of textual mapping in EU law.” In The Ashgate Handbook of Legal Translation, edited by Le Cheng, King Kui Sin & Anne Wagner, 177–192. ), which partially deviates from mainstream postulates of phraseological studies (see Corpas Pastor 2017Corpas Pastor, Gloria 2017 “Collocations in E-Bilingual Dictionaries: From Underlying Theoretical Assumptions to Practical Lexicography and Translation Issues.” In Collocations and Other Lexical Combinations in Spanish. Theoretical and Applied Approaches, edited by Sergi Torner Castells and Elisenda Bernal, 139–160. London: Routledge. for an overview). are collocates of terms which embed terms in cognitive scripts and the text, evidencing combinatory properties of terms. N + V term-embedding collocations can be deemed prototypical in this category and provide important conceptual domain information. They denote what one can typically do with (or to) the object denoted by the base noun, e.g.: The EU ‎Commission launched an infringement procedure ; La Comisión es consciente de las preocupaciones que señalan los peticionarios [The Commission is aware of the concerns raised by the petitioners].

• Finally, lexical collocations are routine formulae at the microstructural level which are not built around terms. They include inter/intratextual referential patterns, such as collocates of editing units, other recurrent patterns referred to as qualifications and non-terminological lexical bundles. In contrast to term-embedding collocations and multi-word terms, recurrence is an important criterion in their identification, e.g.: according to the Water Framework Directive ; propuesta de resolución conforme al artículo 227 punto 2 [motion for a resolution pursuant to Rule 227 (2)].

Term-forming patterns and lexical collocations are outside the scope of this paper on verbal patterns with NEs as subjects or complements.

3.2Choice of corpus

For this study we will use an enlarged version of the PETIMOD corpus (described in Corpas Pastor and Sánchez Rodas, 2022). PETIMOD is a parallel intermodal corpus composed of citizens’ petitions and other documents related to the European Parliament’s Committee on Petitions (PETI). It comprises two subcorpora: (a) original texts and speeches in English and Spanish (PETIMOD_ORIG), and (b) their corresponding translations into Spanish and interpretations into English (PETIMOD_MEDIATED). The genres included are basically notices to members, speeches by MEPs and speakers invited to the Committee of Petitions’ sessions, and non-petitional public documents discussed in the sessions (e.g. reports, opinions).

3.2.1Corpus size

The initial version of the corpus (PETIMOD 1.0) consisted of all petitions discussed during the sessions of February 2020, and all original Spanish (ES) speeches and their English (EN) interpretations of the session of 19th February 2020. Corpus compilation involved a collection of spoken and written documents. Data collection of written documents was rather straightforward: through the eMeeting portal33. https://​emeeting​.europarl​.europa​.eu​/emeeting​/committee​/agenda​/202002​/PETI​?meeting​=PETI​-2020​-0219​_1P​&session​=02​-19​-10​-00. of the European Parliament notices to members, reports and opinions were accessed, downloaded, reformatted (from .pdf to .txt), coded and stored. In comparison, the collection of oral data and transcriptions was more complex and time-consuming. It involved: (1) accessing oral data via the Webstreaming section of the EP Committees site; (2) downloading and storing recordings in HQ .mp4 format; (3) carrying out automatic speech recognition (ASR) plus automatic text transcription (ATT) of audiofiles with YouTube; (4) manually checking and revising in line with EPTIC conventions (Bernardini et al. 2018Bernardini, Silvia, Adriano Ferraresi, Mariachiara Russo, Camille Collard, and Bart Defrancq 2018 “Building Interpreting and Intermodal Corpora: A How-to for a Formidable Task.” In Making Way in Corpus-based Interpreting Studies, edited by Mariachiara Russo, Claudio Bendazzoli & Bart Defrancq, 21–42. Singapore: Springer. ) and the EU Interinstitutional Style Guide (European Union 2012aEuropean Union 2012a Interinstitutional Style Guide 2011. Luxembourg: Publications Office of the European Union. and 2012bEuropean Union 2012b Libro de estilo interinstitucional 2011. Luxembourg: Publications Office of the European Union. ),44.The transcriptions employed the English and Spanish ISG PDF editions of the year 2012, together with the latest modifications of the ISG website up to the submission date, that is, March 2021 (see the news archive at http://​publications​.europa​.eu​/code​/en​/en​-000300​.htm). However, a new PDF edition for all EU languages was released in 2022 (https://​data​.europa​.eu​/doi​/10​.2830​/215072). among others; and (5) coding and storing.

For this study, the corpus has been enlarged to include the English and Spanish draft agendas55. https://​emeeting​.europarl​.europa​.eu​/emeeting​/committee​/en​/agenda​/202002​/PETI​?meeting​=PETI​-2020​-0219​_1P​&session​=02​-19​-10​-00. and minutes66. https://​emeeting​.europarl​.europa​.eu​/emeeting​/committee​/en​/agenda​/202004​/PETI. from the February 2020 meeting, as well as the original Spanish speeches from the 20th February session and their interpretations into English. This session was chosen because it had similar features to that of 19th February already compiled (original Spanish speeches on related environmental topics). Draft agendas and minutes are very relevant documents in the context of committee meetings which were previously used for guidance at the first steps of PETIMOD compilation, especially for identifying and codifying the speakers’ interventions. This time they were also included in the corpus itself. In an NER-oriented study, these two text genres can be very fruitful, as they provide additional NE samples apart from the people and topics intervening in each session. Section B of draft agendas, for example, lists petitions which are proposed for closure in the light of the Commission’s written reply or other documents received. In accordance with the committee’s Guidelines, these items are not discussed during the meeting, but any PETI Member may ask before the end of the meeting for an item in section B to be kept open (European Parliament 2018European Parliament 2018 Guidelines: Committee on Petitions. https://​www​.europarl​.europa​.eu​/cmsdata​/138889​/1145997EN​.pdf). Minutes also include additional information in the form of requests on certain petitions not necessarily reflected in the agenda, and an attendance list of each meeting with MEPs, commissioners, guests, journalists, etc. Table 1 summarizes the size of PETIMOD 2.0 (in total, per component and per language). The total number of documents, running words (tokens) and word types (types) have been calculated using ReCor.77.ReCor is a solution to determine the minimum size of a corpus or a textual collection, regardless of language or textual genre of the collection, establishing therefore the minimum threshold for representation by an algorithm (N-Cor) and analyzing lexical density according to the incremental increase in the corpus (http://​www​.lexytrad​.es​/en​/resources​/recor​-3/). The corpus size was increased by 86 documents, 3,544 types and 23,608 tokens.

Table 1.PETIMOD 2.0

PETIMOD 2.0	Documents	Types	Tokens
PETIMOD2_ORIG_EN	21	5,330	52,421
PETIMOD2_ORIG_ES	81	3,072	18,409
PETIMOD2_MEDIATED_EN	81	2,025	15,709
PETIMOD2_MEDIATED_ES	21	6,262	61,377
PETIMOD2_EN (ORIG + mediated)	102	7,355	68,130
PETIMOD2_ES (ORIG + mediated)	102	9,334	79,786
PETIMOD2	204	16,689	147,916

3.2.2Transcription conventions and revisions

The initial set of transcription conventions were based on Bernardini et al. (2018Bernardini, Silvia, Adriano Ferraresi, Mariachiara Russo, Camille Collard, and Bart Defrancq 2018 “Building Interpreting and Intermodal Corpora: A How-to for a Formidable Task.” In Making Way in Corpus-based Interpreting Studies, edited by Mariachiara Russo, Claudio Bendazzoli & Bart Defrancq, 21–42. Singapore: Springer. , 26–27). For this study, features and codes used in the first compilation process were also revised. The purpose was to better accommodate transcription conventions to our NER-based methodology, removing oral features which could hinder the recognition of named entities or complex phraseology, a task proved problematic in our previous research (cf. Corpas Pastor and Sánchez Rodas, 2022Corpas Pastor, Gloria & Fernando Sánchez Rodas 2022 NLP-enhanced Shift Analysis of Named Entities in an English Spanish Intermodal Corpus of European Petitions. In Marta Kajzer-Wietrzny, Adriano Ferraresi, Ilmari Ivaska & Silvia Bernardini (eds.), Mediated discourse at the European Parliament: Empirical investigations, 219–251. Berlin: Language Science Press. https://​langsci​-press​.org​/catalog​/book​/343). Thus, from the initial set a series of non-relevant or problematic items were removed, namely filled pauses (ehm), mid-word pauses (proposal /pro_posal), non-verbalized noises ([applause]), non-standard pronunciation (sun /su:n/) and truncated words (admin-). Our proposal for transcription conventions also includes a new feature (borrowings/denominations), as illustrated in Table 2.

Our transcription method follows a minimalistic line, ignoring the representation of purely oral features such as filled and mid-word pauses, non-verbalized noises, non-standard pronunciation, and truncated words. Mispronunciations are relegated to lapses affecting NEs. In these cases, the NE is corrected in transcription and the complete mistaken form written between single slashes (/) right after; this allows NE recognition for each case without compromising error analysis later on. Single-word repetitions are not conveyed, and neither are clarifications from interpreters. On the other hand, more flexibility is given to the use of ellipses (…). Apart from silent pauses, ellipses are also used for signalling relevant reformulations (that is, when an entire word or phrase is uttered and then completely amended) and sub-sentence segments (e.g. appositions or additional discourse markers at the beginning or the middle of a sentence). An inclusion of our own is the use of quotation marks to identify borrowings and denominations in plain-text format.88.English quotation marks are used for both languages.

Table 2.Simplified transcription conventions

Feature	Code
Silent pause / reformulations / sub-sentence segments	…
Rising intonation	?
Inaudible segment	#
Mispronunciation	Parlamento /parlo’mento/
Ambiguity	NA
Overlapping talk	NA
Sentence-like segments99.Double slashes (//), used in EPIC, were preferred over the original full stop (.) from the start because of aesthetic reasons.	//
Borrowings/denominations	“fitness check” de las Directivas

As in the first version of the corpus, spelling and capitalization were revised using a number of related resources for each language. This time, however, the process was not two-phased: on the contrary, segmentations and spelling revision could be performed simultaneously, taking advantage of our simplified approach to transcription. The English1010. https://​publications​.europa​.eu​/code​/en​/en​-000100​.htm. and Spanish1111. https://​publications​.europa​.eu​/code​/es​/es​-000100​.htm. versions of the Interinstitutional Style Guide (ISG) were used as reference, from which a selection of additional spelling resources stemmed.1212.These resources are listed as “Reference Works” for the English publications in the Official Journal and “Obras de consulta” for Spanish-specific conventions, respectively. In those cases in which language offers different correct usages, we follow the indications of the Spanish version of the ISG (European Union 2012b, 151), which states that the prevailing criteria will be those of EU translation services (e.g. for capitalization) and the Publications Office agreed norms for all languages (e.g. for acronyms). The PETIMOD corpus (version 1.0.) was also consulted throughout the revision process for the sake of coherence.

3.3Named entity recognition

Parallel to the enlargement of the corpus and simplification of the transcription conventions, the named entity recognition (NER) procedure was also revised. In Natural Language Processing (NLP), NER is a subtask of information retrieval consisting in the location and classification of unambiguous objects or items encountered in a document or in a corpus (for a general overview, see Nasar, Jaffry, and Malik 2021Nasar, Zara, Syed Waqar Jaffry, and Muhammad Kamran Malik 2021 “Named Entity Recognition and Relation Extraction: State-of-the-Art.” ACM Computing Surveys 54(1), 1–39. and Nouvel, Ehrmann, and Rosset 2016Nouvel, Damien, Maud Ehrmann, and Sophie Rosset eds. 2016 Named Entities for Computational Linguistics. Hoboken, NJ: John Wiley & Sons, Inc. ). Examples of real-world “objects” or NEs found in our corpus are Czech Republic, European Social Fund, Angel Dzhambazki, the Greens, the Ombudsman for Persons with Disabilities, etc.

3.3.1Extraction of entities and system performance

In Corpas Pastor and Sánchez Rodas (2022)Corpas Pastor, Gloria & Fernando Sánchez Rodas 2022 NLP-enhanced Shift Analysis of Named Entities in an English Spanish Intermodal Corpus of European Petitions. In Marta Kajzer-Wietrzny, Adriano Ferraresi, Ilmari Ivaska & Silvia Bernardini (eds.), Mediated discourse at the European Parliament: Empirical investigations, 219–251. Berlin: Language Science Press. https://​langsci​-press​.org​/catalog​/book​/343, the chunking, detection and extraction of NEs were carried out in two phases. First, entities were retrieved automatically with the VIP1313.For a brief description of the VIP system, see Corpas Pastor (2021)Corpas Pastor, Gloria 2021 “Technology Solutions for Interpreters: The VIP System.” Hermēneus. Revista de Traducción e Interpretación 23:91–123.. NER module and exported to an Excel file. The NER module integrated SpaCy1414.SpaCy is a free open-source library in Python (https://​spacy​.io/). and pre-trained models for English and Spanish.1515.The VIP NER annotation schemes distinguish at least four basic entity types: named persons or families (PER, e.g. Dorthe Christensen, Ádám Kósa); names of politically or geographically defined locations (LOC, e.g. Mar Menor Coastal Lagoon, Salinas y Arenales de San Pedro del Pinatar), names of corporate, governmental or other organizational entities (ORG, e.g. EU, Tribunal de Cuentas Europeo) and miscellaneous entities (MISC, e.g. Tihange 2, AAE UE-Japón). In order to assess the system performance (and, therefore, the accuracy of results), precision and recall were calculated.1616.Precision refers to the fraction of relevant NEs (i.e. a total number of correctly retrieved NEs minus errors) among all retrieved instances whereas recall refers to the fraction of retrieved instances among all relevant instances found in the corpus. For precision, two levels of analysis were established: (1) relevant NEs, i.e. entities that had been correctly extracted, and (2) relevant NEs that had been correctly ascribed to four categories: person (PER), organization (ORG), location (LOC) and miscellaneous (MISC). Precision results for English were 0.697/0.603; and 0.513/0.385 for Spanish. These results showed a better performance for English, especially when relevant and correctly classified NEs were considered. During the second phase of the analysis, the list of automatically retrieved NEs was supplemented with entities extracted manually with SketchEngine.1717. https://​www​.sketchengine​.eu/. Recall of our NER system was then calculated (relevant NEs / relevant and correctly classified NEs): 0.773/0.969 for English, and 0.677/0.612 for Spanish.

Low precision/recall results were attributed to various issues related to transcription conventions, lack of finer-grained named entity categories, or different pre-trained language models. In order to overcome those limitations, transcription conventions were revised and simplified (cf. 3.2). In addition, we adopted a new finer-grained classification of NEs by replacing the initial four categories with a longer list of pretrained categories that can accommodate up to 18 entities for English and 14 for Spanish (people, organizations, places, money, time, date, laws, languages, products, numbers, etc.).

For NER we used DeepPavlov,1818. https://​deeppavlov​.ai/. an open-source framework for deep learning tasks in Python (Burtsev et al. 2018Burtsev, Mikhail, Alexander Seliverstov, Rafael Airapetyan, Mikhail Arkhipov, Dilyara Baymurzina, Nickolay Bushkov, and Marat Zaynutdinov 2018 “Deeppavlov: Open-source library for dialogue systems.” In Proceedings of ACL 2018, System Demonstrations, 122–127. Melbourne: Association for Computational Linguistics. ). The 18 categories have been trained on OntoNotes 5.0 (Weischedel et al. 2013Weischedel, Ralph, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, and Ann Houston 2013 “OntoNotes Release 5.0 LDC2013T19.” Philadelphia: Linguistic Data Consortium. ) and BERT word embeddings have been used (Devlin et al. 2019Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova 2019 “Bert: Pre-training of deep bidirectional transformers for language understanding.” https://​arxiv​.org​/abs​/1810​.04805v2). Unlike SpaCy-based models, which employ convolutional neural networks, the BERT model offers improved performance for almost all NLP tasks. Table 3 compares performance of SpaCy and DeepPavlov NER models on PETIMOD 2.0.

Table 3.SpaCy and DeepPavlov performance compared (precision)

PETIMOD 2.0 (NER)	SpaCy			DeepPavlov
	EN	ES		EN	ES
errors	322	820		309	383
relevant nes retrieved	2,256	1,261		1,616	2,185
precision	0.875	0.605		0.839	0.851
	0.74			0.87

Precision results with SpaCy improve slightly (+0.178 for English, and +0.002 for Spanish; mean precision rate: 0.74; difference rate among languages: 0.27). This could be due to the NER-friendly simplified transcription conventions used in PETIMOD 2.0. By contrast, NER based on DeepPavlov shows slightly lower precision results for English (−0.036), but it exhibits better performance for Spanish (+0.246), with a higher mean precision rate for both languages (0.87) and a smaller difference rate (0.012).

3.3.2Phraseological Pattern Extraction for NEs

The VIP Query Corpus (QC) module and the NER module have been used to study formulaicity. First, the NER module has been used to extract NEs for manual pattern extraction. Secondly, the QC Pattern functionality has been used to combine part of speech (PoS) tags with the named entity (ENT) tag for automatic pattern extraction. Thirdly, the QC concordance functionality was also used to perform manual KWIC searches. Since the QC Pattern functionality is programmed in SpaCy, it was necessary to complement the automated retrieval with manual KWIC searches in order to locate phraseological patterns for the remaining NEs that had only been recognized with DeepPavlov. This was also necessary at times in order to retrieve contextualized examples of the patterns from the corpus. Targeted searches using key verbs and/or NEs were performed in order to retrieve candidate patterns and/or NEs spotted in the texts as well, which had not been identified neither by DeepPavlov nor SpaCy. Finally, results were imported automatically to Excel files or entered manually when needed.

Our analysis focuses on verbal patterns, divided into (i) entity-as-subject patterns and (ii) entity-as-complement patterns. Verbal patterns convey formulaicity at the level of text organization, grammatical relations, and collocational networks. Table 4 shows the list of all phraseological patterns used to query the PETIMOD 2.0 corpus.

Table 4.List of total verbal patterns searched for in the corpus

verbal patterns
Entity-as-subject patterns	Entity-as-complement patterns
ENT + V	V + ENT
ENT + * + V	V + * ENT
ENT + V + *	V + * + ENT
	V + ENT + *
	V + * + * + ENT
	V + * + * + * + ENT

4.1Distribution of NEs

Table 5 shows the general distribution of NEs across subcorpora. Figures for absolute and normalized frequencies are provided. The most represented categories in non-mediated English are PER (4.17), DATE (3.99), ORG (3.75), LAW (3.37) and MISC (2.18). However, translations into Spanish reflect a slightly different distribution as regards those main categories: DATE (5.46), ORG (4.06), LAW (3.46), PER (3.43), and MISC (3.18). These differences could be due to the impact of non-translated Spanish (as DATE is higher in the normalized frequency rank, while ORG is 0.1 more frequent than PER), or else be a direct consequence of various translation shifts. The fact that almost all main categories (except PER) present more instances in translated Spanish than in their original English texts also points to the existence of explicitation. In any case, the lower number of instances of PER found in translated Spanish is not in contradiction to explicitation, as original PER might have been translated as ORG.

Regarding interpreted English, the actual figures seem to point to simplification instead, as this subcorpus systematically includes fewer NEs than the corresponding original Spanish speeches. For instance, there are 36 instances of PER in original Spanish as opposed to 17 in interpreted English (the same applies to most categories). The lower number of dates (15) and cardinals (14) in the interpreted corpus compared to the higher number of instances found in the Spanish original speeches (43 and 36, respectively) could also be a sign of cognitive overload or difficulties experienced by the interpreters. Total number of NEs seem to support those two opposed general tendencies: the total number of NEs in the translated Spanish subcorpus (1,631) is higher than in the original English texts (1,239), whereas interpreted English contains fewer NEs (145) than original speeches in Spanish (294).

4.2Text-organizing patterns

Text-organizing patterns (Table 6) are the most common category in ORIG_EN (94/139 subpatterns, 68%), and also the most varied one (11/24, 46%). 9/11 variations (82%) are built around entities in subject position. The most frequent associated NE types are DATE (42/139, 30%), ORG (36/139, 26%) and TIME (3/139, 2%). This is an indication of the importance of time organization in the activity of the Committee on Petitions, which is reflected in the different parts dividing the sessions (draft agendas and minutes), but also in the administrative processes summarized in the petition themselves. Another common point is that practically all subpatterns (10/11, 91%) contain a past participle or past continuous form. This could indicate a certain delay between the time when petitions are received and the moment at which the Committee or the Commission takes action. Each pattern can also be related to specific text-organizing functions. Two out of three ENT + V + * subpatterns, for example, are associated with concluding remarks in the petition summaries, whereas the entity-as-complement subpatterns reflect more practical, administrative functions such as deadlines, committee meeting pauses and starts, etc.

When translating English sequences with text-organizing patterns into Spanish (Table 7), translators depart considerably from non-mediated Spanish. Most patterns in translated Spanish present entities as complements (6/8, 75%) and the most frequent NE types are PER and DATE instead of ORG (cf. Table 4). This reveals that although translated texts generally transfer the usage of text-organizing patterns for the same organizing functions (deadlines, concluding remarks, etc.), normalization seems to operate simultaneously on a smaller scale, affecting syntax (and interestingly also NE types) by accommodation to the Spanish norms. Consider the subpattern presentada + * + ENT, which does not exist in ORIG_EN. Non-translated petitions opt for a sequence that omits any verbal form (e.g. Petition No 1106/2018 by Alexander ‎Edberg Thorén). The shift in translation (Petición n.º 1106/2018, presentada por Alexander Edberg Thorén) adheres better to the target language norms, which demand the introduction of a past participle before the preposition por.

Another difference with ORIG_ES is that these past forms (5/11 patterns, 45%) are not completely hegemonic, but coexist with present simple (2/11 patterns, 18%) and future simple forms (1/11 pattern, 9%). This could also be compatible with normalization, evidenced by a decreased use of passive forms in Spanish and some differences in textual conventions between languages (see minutes, where resumed + * + ENT shifts to reanuda + * + * + ENT). Finally, explicitation is also observed through translation shifts. For instance, the subpattern apruebe + * + ENT typically occurs in MED_ES. It refers to a future process in which a second EP committee would approve (or not) a motion of resolution helped by the opinion of the Committee on Petitions (incorpore las siguientes sugerencias la propuesta de resolución que apruebe). The non-translated version is less explicit and does not make reference to the autonomy of the neighbouring committee (incorporate the following suggestions into its motion for a resolution).

Text-organizing patterns in the non-interpreted Spanish speeches (Table 8) are strongly associated with PER entities (11/22). Most of the ritualized, repeated forms are used either to present the initiators of the petitions (presentada por Olga Daskali) or to give the floor to them. In the first case, it is worth noting that the oral pattern seems to borrow the above-mentioned phraseme appearing in the title of Spanish petitions (presentada por). In the second case, the complex structure tiene la palabra is used not only to give the floor to certain participants, but also to remind them of the intervention time available, as in the example tiene la palabra por cinco minutos. This introduction, however, seems to have more frequent associations with simpler patterns, such as V + ENT (tiene cinco minutos).

When interpreted into English, text-organizing patterns (Table 9) occurs less frequently (34%) than in original subcorpora of Spanish (55%) and English (68%). Variation is also smaller in mediated English when compared to both originals (ORIG_EN 46%, ORIG_ES 33%, MED_EN 31%), although in the case of ORIG_ES there is almost no difference. These figures point towards simplification in the interpreted discourse, which can be additionally supported by two shifts. The interpreted concordance you have five minutes // tell us about your petition has a simpler syntax than the non-interpreted Spanish tiene cinco minutos para plantear su intervención). Another example of simplification is petition 0827/2008 presented by Olga Daskali // Greek //, in which the original Spanish de nacionalidad griega is condensed in only one word. Negative transfer (i.e. interference) can also be observed in the use of the subpattern presented + * + ENT.

4.3Grammatical patterns

Grammatical patterns (Table 10) are the second most frequent category in ORIG_EN (28/139, 20%). They are highly associated with ORG NEs (20/28, 71%) and mostly feature entity-as-subject patterns (4/6, 67%). Deontic modal verbs represent most findings in this category,2020.It must also be noted that different modal verbs appear in different types of patterns. Consider for example will, which was previously associated to text-organizing patterns (see Table 10). although important differences between them should be established. Weaker obligation patterns, such as should (7/28) and may (3/28), are more numerous than the stronger shall (6/28) and must (3/28). As an example, the opinion PETI should not draft is weaker than the sentence The European Parliament shall draw up a proposal, which is in fact a quotation from the EC Treaty. Two impersonal structures were also detected, with different forms: one is passive (provided + V + ENT) and the other one is a third-person-singular present tense (calls on the Commission to…).

Grammatical patterns in MED_ES (Table 11) are more numerous than in ORIG_EN (27% vs. 20%), which could be an indicator of aggregate transfer. Another indicator of transfer would be the stability of the relative percentage of entity-as-subject patterns, which remains practically the same as in ORIG_EN (25% vs. 29%).

Together with transfer, normalisation is hypothesized at the microtextual level on the basis of three shifts. Deontic modal verbs are not always calqued into Spanish, even when possible, which reflects the preference of the target language for alternative forms of expressing modality. This is the case for two example patterns which share the same verb (elaborar).2121.This may be regarded as an additional hint of simplification and/or convergence. In the first, Coordinators decided that PETI should not draft changes to Los coordinadores deciden que la Comisión PETI no elabore (subjunctive).2222.Also, the use of subjunctive with the verb decidir conveys a higher degree of assertiveness than the alternatives debería elaborar and debe elaborar. In the second example, The European Parliament shall draw up a proposal changes to el Parlamento Europeo elaborará un proyecto (future simple). Even when equivalent periphrastic structures are used, involuntary errors occur due to the influence of target language rules. In the example the Commission may refer the case to the European Nuclear Safety Regulators Group → la Comisión puede remitir el asunto al Tribunal de Justicia de la Unión Europea, the second NE is a human mistake, probably caused by the usual association of the term case with the Court of Justice of the European Union.

Grammatical patterns associated with entities in the non-mediated Spanish subcorpus (Table 12) are less numerous than in the mediated subcorpus. Occurrence percentages of grammatical patterns grow steadily from ORIG_EN (20%) and MED_ES (26%) to ORIG_ES (28%). As in non-translated English (71%) and translated Spanish (88%), NEs are mostly ORG (82%). Similarly to text-organizing patterns, these progressively increased percentages can be read as a sign of transposition. Additional evidence might be that 3/5 verbs were already present in MED_ES (pedir, facilitar, deber) and that such patterns present two subtle shifts apparently caused by a remote, indirect influence of the source language. Consider the NE associated with the pattern pedimos + * + ENT (Comité PETI), literally borrowed from PETI Committee, and the subpattern pedir + * + ENT, a grammatical mistake in standard Spanish possibly influenced by English.

In the interpreted English subcorpus, two new grammatical patterns (Table 13) are introduced. The first one is ask + * + ENT, which functions as a register-down alternative to call + * + ENT. In fact, it presents more occurrences (nine against six), which will again confirm the informality of the interpreted English subcorpus. Another new pattern is ENT + would + *. The PER entity functioning as subject (Sánchez) is an informal alternative to Pedro Sánchez. In a further example, the subpattern ENT + should + * reveals another register-down shift which additionally decreases the number of words in the interpreted corpus (el Mar Menor debe ser una cuestión de Estado → Mar Menor should be tackled by the State). As it happened in MED_ES, shifts in modal verbs imply shifts of meaning and register also in the main verb.

4.4Term-embedding collocations

Term-embedding collocations display some similar numbers in all subcorpora because prototypical NEs (Commission/European Commission) and similar patterns (N + V) were searched for to offer an exemplification. 100% of the total occurrences are associated with ORG NEs and relate to a subject entity. The rest of the indicators, however, show certain differences between components (see below). Although these numbers may be biased by our conscious choice, the findings might also suggest transfer in translated and interpreted subcorpora.

This category is least frequent (17/139, 12%) in non-translated English (Table 14). Even though the subpatterns are distributed among ENT + V (43%) and ENT + V + * (57%), personification is a common feature of all term-embedding collocations found in ORIG_EN. The European Commission is construed as a human being with the ability of performing both mental and physical tasks. The first group (12/17, 71%) describes the often diplomatic postures maintained by this body in relation to the citizens’ petitions (Commission + is + aware, Commission + considers + that, Commission + would + like, Commission + maintains, Commission + understands).2323.A further note on these patterns would be their similarity with other categories. The example The Commission would like to reiterate contains a grammatical pattern, but it conveys a specific idiomatic meaning (“The Commission sends a friendly reminder”). Conversely, the patterns Commission + maintains and Commission + understands could be also classified as text-organizing, given that they are embedded in the concluding section of petitions. In the second group (5/17, 29%), the auxiliary can is introduced as a mechanism to limit more direct actions available for this entity. The main verb take can be spotted in both examples (The Commission cannot take on the role… and The Commission can take legal action…).

Collocates of Commission in translated Spanish (Table 15) are lower in numbers compared to the non-translated English corpus (7% vs. 12%), and equally reflect personification. Nevertheless, some patterns are redistributed because of alternative verb choices. Comisión + entiende, for example, presents one more hit because the verb entender has been used for translating both understand and consider. Similarly, the pattern Commission + would + like is split into two (Comisión + desea and Comisión + querría). The example La Comisión desea reiterar que… is considerably less distant than The Commission would like to reiterate, adding an affective nuance not found in the English personifications.

Non-mediated Spanish term-embedding collocations (Table 16) are more frequent (17%) than in the corresponding mediated subcorpus (7%), which supports the hypothesis of simplification of mediated discourse. Other translationese traits are reflected in the percentage of subpatterns. Only 29% (two out of seven) are shared between original and mediated Spanish. Despite a more reduced number of occurrences (7), collocations also reflect personification phenomena in the non-interpreted Spanish corpus. Indeed, metaphors acquire more diverse forms than in the written corpora. The affective verb desear is repeated in this subcorpus and is linked to two perception verbs which embody (and thus personify) the institution (la Comisión desea siempre escuchar a todas las partes… conocer su visión //). Additionally, new affective meanings are introduced (la Comisión comparte la preocupación…), together with verbs of speech (la Comisión Europea afirma…) and others by which the institution performs tasks humans would (la Comisión Europea incoó un procedimiento… and la Comisión Europea publicó…). This could be a hedging mechanism, designed to hide the real human subject of such procedures. As can be observed, ENT + V patterns with Comisión Europea can be found in this subcorpus, but not in the written collections. It is possible that this terminological variant is introduced in non-interpreted Spanish for better distinguishing the executive Commission and the PETI Committee.

In the interpreted English subcorpus (Table 17), the total number of term-embedding collocations (9) remains quite similar to non-interpreted Spanish. Even though a new mental verb is introduced (believe), the rest of the examples are associated with more physical meanings, even with an idea of movement and urgency which cannot be found in the non-interpreted Spanish (the Commission needs to inform us, measures the Commission takes, the Parliament but also the Commission act to find a solution…).

Finally, these data are in line with our previous findings (Corpas Pastor and Sánchez Rodas, 2022Corpas Pastor, Gloria & Fernando Sánchez Rodas 2022 NLP-enhanced Shift Analysis of Named Entities in an English Spanish Intermodal Corpus of European Petitions. In Marta Kajzer-Wietrzny, Adriano Ferraresi, Ilmari Ivaska & Silvia Bernardini (eds.), Mediated discourse at the European Parliament: Empirical investigations, 219–251. Berlin: Language Science Press. https://​langsci​-press​.org​/catalog​/book​/343). Simplification appears to be a contingent feature which depends on the mediation mode and the source languages involved (and also on the topic of the source text). Another relevant finding of this study is that there are clear differences in shifts between EN-ES translations and ES-EN interpretations of NEs in the Petitions Committee, as well as normalisation of specialized phraseology, especially in interpreted English (from Spanish).